A feature that is part of the feature subset used by
a learning algorithm is a good feature to use if it is
either a good predictor of the class by itself, or a good
predictor of the class when taken together with some
other subset of features in the set. At the same time,
it should not be redundant given the other features in
the selected feature set. Langley (1994) notes that feature
selection algorithms that search through the space
of feature subsets must address four main issues: the
starting point of the search, the organization of the
search, the evaluation of feature subsets and the criterion
used to terminate the search. Different algorithms
address these issues differently.