Optimal Extension of Error Correcting Output Codes

Error correcting output codes (ECOC) represent a successful extension of binary classifiers to address the multiclass problem. In this paper, we propose a novel technique called ECOC-ONE (Optimal Node Embedding) to improve an initial ECOC configuration defining a strategy to create new dichotomies and improve optimally the performance. The process of searching for new dichotomies is guided by the confusion matrices over two exclusive training subsets. A weighted methodology is proposed to take into account the different relevance between dichotomies. We validate our extension technique on well-known UCI databases. The results show significant improvement to the traditional coding techniques with far few extra cost.