Classiers use a multitude of techniques to decide, whether an unknown instance, described by a vector of
feature values, belongs to a certain class. However most of them base their decisions somehow on the fact
that the more similar1
a novel instance is to a set of instances with known class membership, the higher
is the probability, that this instance is in the same class. The similarity between instances is therefore a
crucial information and it is clear, that instance-representations, leading to a reinforcement of similarities
between instances in the same and deterioration of similarities between instances in dierent classes can
lead to an improvement of classier accuracy