Two stage reject rule for ECOC classification systems

Abstract

The original task of a multiclass classification problem can be decomposed using Error Correcting Output Coding in several two-class problems which can be solved with dichotomizers. A reject rule can be set on the classification system to improve the reliability of decision through an external threshold on the decoding outcomes before the decision is taken. If a loss-based decoding rule is used, more can be done to make such external scheme works better introducing a further reject stage in the system. This internal approach is meant to single out unreliable decisions for each classifier in order to proficiently exploit the properties of loss decoding techniques for ECOC as proved by experimental results on popular benchmarks. © 2011 Springer-Verlag.