On ECOC as binary ensemble classifiers

Abstract

The Error-Correcting Output Codes (ECOC) is a representative approach of the binary ensemble classifiers for solving multi-class problems. There have been so many researches on an output coding method built on an ECOC foundation. In this paper, we revisit representative conventional ECOC methods in an overlapped learning viewpoint. For this purpose, we propose new OPC based output coding methods in the ECOC point of view, and define a new measure to describe their properties. From the experiment on a face recognition domain, we investigate whether a problem complexity is more important than the overlapped learning or an error correction concept. © Springer-Verlag Berlin Heidelberg 2005.