ECOC-ONE: A Novel Coding and Decoding Strategy

Error correcting output codes (ECOC) represent a classification technique that allows a successful extension of binary classifiers to address the multiclass problem. In this paper, we propose a novel technique called ECOC-ONE to improve an initial ECOC configuration by including new dichotomies guided by the confusion matrix over exclusive training subsets. In this way, the initial coding represented by an optimal decision tree is extended adding binary classifiers forming a network. Since not all dichotomies have the same relevance, a weighted methodology is included. Moreover, to decode we introduce a new distance to attenuate the error accumulated by zeros in the ECOC-ONE matrix. We compare our strategy to other well-known ECOC coding strategies on the UCI data set achieving very promising results