Aggregation Ordering in Bagging

G. Martínex-Muñoz and A. Suárez (Spain)


Machine learning, bagging, ensemble pruning, decision trees


The order in which classifiers are aggregated in ensemble methods can be an important tool in the identification of subsets of classifiers that, when combined, perform better than the whole ensemble. Ensembles with randomly or dered classifiers usually exhibit a generalization error that decreases as the number of classifiers that are aggregated increases. If an appropriate order for aggregation is chosen, the generalization error reaches, at intermediate numbers of classifiers, a minimum, which lies below the asymptotic error of the ensemble. This work presents some heuris tics that exploit the relations between the classifiers in a bagging ensemble to identify the appropriate ordering and then select a subset for aggregation according to a desired amount pruning. The resulting subensembles are smaller and improve the classification performance of the original ensemble.

Important Links:

Go Back