We have developed a family of feature weighting techniques that vary in the
degree of locality with which the feature weights are calculated. We present
results of tests showing that at least one of the CDW variants signicantly
improves classication accuracy for nine of eleven benchmark classication tasks.
Because no single technique proved to be the best in every task, we conclude
that dierent tasks require diering degrees of locality in feature weighting. This
justies the use of a family of techniques, and suggests that some pre-testing
using cross-validation on a particular task is necessary in order to determine the
amount of locality required.