Bagging in the Presence of Outliers

We investigate the performance of bagging methods in the presence of outliers. The results are best illustrated and intuitively explained for classical classification problems to which we shall restrict our focus in this paper. It is shown that bagging methods can improve the resistance of classification rules to outlier contamination, especially if an m-out-of-n bagging scheme is used. However, the outlier-reduction property does not improve performance to such an extent that outliers are no longer noticeable. It is also shown that in the absence of contamination by outliers the effects of bagging are negligible. Therefore, when bagging is not really needed, deleterious effects that result from employing it are quite small. 1 Introduction The method of bootstrap aggregation, or bagging, was introduced by Breiman (1996a). Recent accounts of its properties, in the settings of both prediction and classification, include those of Breiman (1996b) and Tibshirani (1996). When used for class..