We introduce a framework for the reuse of knowledge from previously tr
ained classifiers to improve performance in a current and possibly rel
ated classification task. The approach used is flexible in the type an
d relevance of reused classifiers and is also scalable. Experiments on
public domain data sets demonstrate the usefulness of this approach w
hen one is faced with very few training examples or very noisy trainin
g data. (C) 1997 Published by Elsevier Science B.V.