next up previous contents
Next: Feature Subset Selection Up: Wrapper Inducers Previous: Discretization filter


Bagging is a wrapper inducer that runs the wrapped inducer, specified in the BAG_INDUCER option, multiple times on subsets of the training set. During classification, the induced classifiers vote and the majority class is chosen [,]. The wrapped inducer must be a regular inducer (not a base inducer).

Bagging seems to work best on unstable inducers, that is, inducers that suffer from high variance because of small perturbations in the data []. Unstable inducers include decision trees ( e.g. , ID3) and perceptrons; an example of a very stable inducer is nearest neighbor, which has a high bias in high-dimensional spaces, but very little variance. What you lose by bagging is the ability to understand the data: you have 20 experts that vote on the label and gains are achieved if they are all good but disagree between themselves many times!

The BAG_REPLICATIONS option determines the number of classifiers to create; the more, the ``better'' in the sense that the result will be more stable. BAG_PROPORTION determines the proportion of the training set that will be passed to each copy of the inducer. The higher the proportion, the larger the internal training set; however, if the data is not perturbed enough, the classifiers won't be different and bagging won't work well []. BAG_UNIF_WEIGHTS is a Boolean option that determines whether the votes are equal or estimated. If the votes are estimated, the portion of the training set that was not used for internal training is used as a test set, and the estimated accuracy is the weight associated with the induced categorizer. Due to high variance in the estimation, this option does not seem to work well in practice.

Ronny Kohavi
Sun Oct 6 23:17:50 PDT 1996