CIFASIS   20631
CENTRO INTERNACIONAL FRANCO ARGENTINO DE CIENCIAS DE LA INFORMACION Y DE SISTEMAS
Unidad Ejecutora - UE
artículos
Título:
Recursive ECOC classification
Autor/es:
TAPIA ELIZABETH; BULACIO PILAR; ANGELONE LAURA
Revista:
PATTERN RECOGNITION LETTERS
Editorial:
ELSEVIER SCIENCE BV
Referencias:
Lugar: Nueva York; Año: 2010 vol. 31 p. 210 - 215
ISSN:
0167-8655
Resumen:
The construction of ECOC (Error Correcting Output Coding) classifiers from one or more constituent ECOC classifiers is proposed. Aiming to boost the accuracy of the overall ECOC system, constituent ECOC classifiers are allowed to Exchange information via shared binary classifiers. A novel decoding algorithm that iteratively combines binary predictions from constituent ECOC classifiers is introduced for this purpose. Aiming to minimize the degrading effects of dependency between binary predictions, the use of sparsely connected ECOC classifiers of small size is recommended. A comprehensive experimental work shows that competitive ECOC classifiers of size at most ⌈3 log2M⌉ can be obtained in this way.