Many feature weighting methods apply weights globally: they use a single weight vector that remains constant throughout testing. However, some domains
contain features that vary in importance across the instance space (Aha and
Goldstone, 1992). Local weighting schemes, where feature weights can vary from
instance to instance or feature value to feature value, may perform better for
such applications. For example, Aha and Goldstone (1992) use a combination of
local and global weights for each training instance, and Hastie and Tibshirani
(1994) use weights produced individually for each test instance. Stanll and
Waltz's (1986) value dierence metric (VDM) takes a slightly dierent approach
by weighting features according to the particular feature values of the test case
and individual training cases. Potentially, local weighting schemes can take into
account any combination of global, test-case, and training-case data. The locality
of particular weighting algorithms can be visualised on a continuum, from global
methods that compute a single weight vector for all cases to extremely local
methods that compute a dierent weight vector for each pair of test and training
cases