To go from the local CDW weights to a global weight
for all features, we average the feature weight vectors across all classes to get
a single global weight vector. This variant can be expected to perform well in
domains where the relevant features are the same for all classes (e.g., LED-24
from the UCI repository (Merz and Murphy, 1996)). The global weights can be
computed by taking a simple mean over all classes or by an average weighted by
the class frequency in the training data. The latter approach gives comparable
overall results, but tends to bias predictions toward the most common classes.
Because recent work has emphasized the importance of minority class predictions
(Fawcett, 1996), we present only results for the simple mean here. We call this
method global mean CDW (GM-CDW)