Evolutionary Ensemble Creation and Thinning

Ensembles are often capable of greater predictive accuracy than any of their individual members. One key attribute of ensembles' success is the notion of diversity. However, the majority voting scheme used in most ensembles treats each classifier as if it contributed equally to the group performance, without capitalizing on the relative improvement offered by each member of the ensemble. Our solution to this problem is to use genetic algorithms to weight the contribution of each classifier. This improves the performance of the ensemble by providing a weighted vote which maximizes collaboration among classifiers. Our approach provides a general-purpose framework for evolutionary ensembles, allowing them to build on top of any collection of classifiers, whether they be heterogeneous or homogeneous. In addition, the ability of our framework to thin ensembles, and its effect on ensemble diversity is presented.