Using additional neighborhood information in a dynamic ensemble selection method : improving the KNORA approach

This work evaluates some strategies to approximate the performance of a dynamic ensemble selection method to the oracle performance of its pool of weak classifiers. For this purpose, different strategies are evaluated to combine the results of the KNORA dynamic ensemble selection method with the results of its built-in KNN used to define the neighborhood of a test pattern during the ensemble creation. The KNN results are considered as additional information which may be combined with the KNORA results to improve the recognition performance. A strong experimental protocol based on more than 60,000 samples of handwriting digits extracted from NIST-SD19 has shown that the fusion of the KNORA results with the results of its built-in KNN is very promising. Keywordsdynamic ensemble selection; oracle; KNORA