A RISK BOUND FOR ENSEMBLE CLASSIFICATION WITH A REJECT OPTION

Signal classification is an important task in numerous application domains that is increasingly being approached through ensemble methods, such as those involving boosting and bootstrap aggregation. In decision support scenarios, it is often of interest for automatic classification algorithms to abstain from making decisions on the most uncertain signals; this is known as classification with a reject option. In this work, a bound on generalization error for ensemble classification with a reject option is derived that involves two intuitive properties of the ensemble: average strength and mean correlation. The bound is shown to be predictive of empirical classification behavior and useful in setting the rejection threshold for a given rejection cost.