An Investigation into the Effect of Ensemble Size and Voting Threshold on the Accuracy of Neural Network Ensembles

If voting is used by an ensemble to classify data, some data points may not be classified, but a higher proportion of those which are classified are classified correctly. This trade off is affected by ensemble size and voting threshold. This paper investigates the effect of ensemble size on the proportions of decisions made and correct decisions. It does this for majority voting and consensus voting on ensembles of neural network classifiers constructed using bagging. It also models the relationships in order to estimate the asymptotic values as the ensemble size increases.