Constructing the shortest ECOC for fast multi-classification

Abstract

Error-correcting output codes (ECOC) is an effective method to perform multi-classification via decomposing a multi-classification problem into many binary classification tasks, and then integrating the outputs of the subtasks into a whole decision. The researches on applying ECOC to multi-classification mainly focus on how to improve the correcting ability of output codes and how to enhance the classification effectiveness of ECOC. This paper addresses a simple but interesting and significant case of ECOC, the shortest ECOC, to perform fast multi-classification at the cost of sacrificing a very small classification precision. The strategy of balancing the positive and negative examples for each binary classifier of ECOC and the method of finding the optimal permutation of all original classes are further given. Preliminary experimental results show, the shortest ECOC uses fewest binary classifiers but can still obtain comparable or close classification precisions with several traditional encoding methods of ECOC. © 2011 Springer-Verlag.