Error-correcting output coding (ECOC) is nowadays an established technique to build polychotomous classification systems by aggregating highly efficient dichotomizers. This approach has exhibited good classification performance and generalization capabilities in many practical applications. In this field much work has been devoted to study new solutions both for the coding and the decoding phase, but little attention has been paid to the algebraic tools typically employed in the Coding Theory, which could provide an ECOC design approach based on robust theoretical foundations. In this paper we propose an ECOC classification system based on Low Density Parity Check (LDPC) Codes, a well known technique in Coding Theory. Such framework is particularly suitable to define an ECOC system that employs dichotomizers provided of a reject option. The experiments on some public data sets have demonstrated that, in this way, the ECOC system can reach good recognition rates when a suitable reject level is imposed to the dichotomizers.
Coding Theory Tools for Improving Recognition Performance in ECOC Systems
MARROCCO, Claudio;SIMEONE, Paolo;TORTORELLA, Francesco
2013-01-01
Abstract
Error-correcting output coding (ECOC) is nowadays an established technique to build polychotomous classification systems by aggregating highly efficient dichotomizers. This approach has exhibited good classification performance and generalization capabilities in many practical applications. In this field much work has been devoted to study new solutions both for the coding and the decoding phase, but little attention has been paid to the algebraic tools typically employed in the Coding Theory, which could provide an ECOC design approach based on robust theoretical foundations. In this paper we propose an ECOC classification system based on Low Density Parity Check (LDPC) Codes, a well known technique in Coding Theory. Such framework is particularly suitable to define an ECOC system that employs dichotomizers provided of a reject option. The experiments on some public data sets have demonstrated that, in this way, the ECOC system can reach good recognition rates when a suitable reject level is imposed to the dichotomizers.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.