Effects of Missing Members on Classifier Ensemble Accuracy

Classification uses a predictive model to predict labels for previously unseen data. Classifier ensembling techniques are used to combine the predictive outputs of multiple classifiers into one final predictive output in order to increase the accuracy of the predictive output. These techniques tend to assume every member classifier will always be available to make a prediction. This paper examines the effects missing members have on the prediction accuracy of a Stacking based ensemble and a Voting based ensemble. To determine the impact, 20 datasets were randomly selected from the UCI Machine Learning Repository; these datasets were used to create multiple Stacking and Voting models. Results indicate that the Stacking ensemble performs poorly in the face of even one missing member, while the Voting ensemble has very little issue with missing members.