Computing and Informatics

Bootstrapped aggregation of classifiers, also referred to as bagging, is a classic meta-classification algorithm. We extend it to a two-stage architecture consisting of an initial voting amongst one-versus-all classifiers or single-class recognizers, and a second stage of one-versus-one classifiers or two-class discriminators used for disambiguation. Since our method constructs an ensemble of elementary classifiers, it lends itself very well to parallelization. We describe a static workload balancing strategy for embarrassingly parallel classifier construction as well as a parallelization of the classification process with the message passing interface. In our experiments, which are evaluated in terms of classification performance and speed-up, we obtained an up to three-fold increase in precision and significantly increased recall values.