Ensemble Selection Using Diversity Networks

An ideal ensemble is composed of base classifiers that perform well and that have minimal overlap in their errors. Eliminating classifiers from an ensemble based on a criterion that reflects poor classification performance and error redundancy with peer classifiers can improve ensemble performance. The Diversity Networks method asymmetrically evaluates each pair of classifiers as a linear combination of individual performance and diversity. This measure is used to prune the ensemble gradually to find a nearly optimal ensemble.