Classifier selection for majority voting

Individual classification models are recently challenged by combined pattern recognition systems, which often show better performance. In such systems the optimal set of classifiers is first selected and then combined by a specific fusion method. For a small number of classifiers optimal ensembles can be found exhaustively, but the burden of exponential complexity of such search limits its practical applicability for larger systems. As a result, simpler search algorithms and/or selection criteria are needed to reduce the complexity. This work provides a revision of the classifier selection methodology and evaluates the practical applicability of diversity measures in the context of combining classifiers by majority voting. A number of search algorithms are proposed and adjusted to work properly with a number of selection criteria including majority voting error and various diversity measures. Extensive experiments carried out with 15 classifiers on 27 datasets indicate inappropriateness of diversity measures used as selection criteria in favour of the direct combiner error based search. Furthermore, the results prompted a novel design of multiple classifier systems in which selection and fusion are recurrently applied to a population of best combinations of classifiers rather than the individual best. The improvement of the generalisation performance of such system is demonstrated experimentally.