We measure the usefulness of a feature for classication by comparing the
distributions of the feature values across various subsets of the training cases.
CDW computes a dierent set of weights for each class in the training set. The weights for a particular class on a given feature are based on a comparison
between the distribution of feature values for the cases in that class and the
distribution of values for cases in all other classes. If the distributions are highly
similar, the feature is considered not useful for distinguishing that class from
others, and it is assigned a low weight. If the distributions are highly dissimilar,
the feature is considered useful, and it is assigned a high weight