A new metric for greedy ensemble pruning

Ensemble pruning is a technique to reduce ensemble size and increase its accuracy by selecting an optimal or suboptimal subset as subensemble for prediction. Many ensemble pruning algorithms via greedy search policy have been recently proposed. The key to the success of these algorithms is to construct an effective metric to supervise the search process. In this paper, we contribute a new metric called DBM for greedy ensemble pruning. This metric is related not only to the diversity of base classifiers, but also to the prediction details of current ensemble. Our experiments show that, compared with greedy ensemble pruning algorithms based on other advanced metrics, DBM based algorithm induces ensembles with much better generalization ability.