“…The idea is to define a set of domain characteristics or meta-features that are relevant for predicting the performance of the given learning algorithms (Aha, 1992;Michie, Spiegelhalter, & Taylor, 1994;Gama & Brazdil, 1995;Brazdil, 1998;Keller, Paterson, & Berrer, 2000;Brazdil, Soares, & Pinto da Costa, 2003). Meta-features may include information concerning error rates of base-learners, so called landmarkers Pfahringer, Bensusan, & Giraud-Carrier 2000) or the structure of induced decision trees (Bensusan, 1998;Peng et al, 2002).…”