Content area
Abstract
Classifier ensembling is a method of combining the decisions of multiple classifiers into one final decision with the expectation that the combined decision will be more accurate than the decision of any one member of the ensemble. Ensembles can also be pruned to create a sub-ensemble with a lower computation time with little to no decrease in accuracy. However, most ensemble generation methods and ensemble pruning methods assume that every member will always be available to provide its input. This research seeks to evaluate the ability of ensembles to handle missing members by using twenty commonly used datasets and seven classifiers ensembled with Majority Voting and Stacking methods. A genetic algorithm adjusted version of Stacking designed to adjust for missing members is also introduced and evaluated. Results show that Stacking based methods perform poorly in the presence of missing members, while Voting methods were largely unaffected by missing members. The genetic algorithm adjusted Stacking typically performed better than regular Stacking in the face of missing members, but still did not perform as well as Majority Vote.
