For some classes, particular feature values may be
more signicant than other values of the same feature. For example, the target
class in the Monks-2 data set is dened by the concept exactly two features have
the value 1." Thus another potentially useful form of local feature weighting allows
the weights to vary locally according to the values of the test instance. A
simple transformation of the case base allows the CDW algorithm to exploit this
form of locality. Specically, the feature set is expanded so that each instance
is described by a set of binary features corresponding to all the feature value
possibilities in the original training set. If instances in the training set T are
described by the set of features F = ff1; :::; fmg, and feature fi takes on values
Vfi = fvfi1; :::; vfiri g across the entire training set, then instances in the transformed
training set T 0 are described by the feature set F 0 = Vf1 [ Vf2 [ ::: [ Vfm