On nearest-neighbor error-correcting output codes with application to all-pairs multiclass support vector machines

A common way of constructing a multiclass classifier is by combining the outputs of several binary ones, according to an error-correcting output code (ECOC) scheme. The combination is typically done via a simple nearest-neighbor rule that finds the class that is closest in some sense to the outputs of the binary classifiers. For these nearest-neighbor ECOCs, we improve existing bounds on the error rate of the multiclass classifier given the average binary distance. The new bounds provide insight into the one-versus-rest and all-pairs matrices, which are compared through experiments with standard datasets. The results also show why elimination (also known as DAGSVM) and Hamming decoding often achieve the same accuracy.