Machine Learning
On Bias, Variance, 0/1—Loss, and the Curse-of-Dimensionality
Data Mining and Knowledge Discovery
An empirical evaluation of bagging and boosting
AAAI'97/IAAI'97 Proceedings of the fourteenth national conference on artificial intelligence and ninth conference on Innovative applications of artificial intelligence
Dynamics of variance reduction in bagging and other techniques based on randomisation
MCS'05 Proceedings of the 6th international conference on Multiple Classifier Systems
Hi-index | 0.00 |
Bagging is a procedure averaging estimators trained on bootstrap samples. Numerous experiments have shown that bagged estimates almost consistently yield better results than the original predictor. It is thus important to understand the reasons for this success, and also for the occasional failures. Several arguments have been given to explain the effectiveness of bagging, among which the original "bagging reduces variance by averaging" is widely accepted. This paper provides experimental evidence supporting another explanation, based on the stabilization provided by spreading the influence of examples. With this viewpoint, bagging is interpreted as a case-weight perturbation technique, and its behavior can be explained when other arguments fail.