A novel decoding strategy based on DS evidence theory

Error-Correcting output codes is an effective decomposing frame to reduce multi-class to binary, in which the research of decoding strategy is especially attracting more and more attention. In this paper, we propose a new decoding approach on DS evidence theory. To achieve this goal, we obtain the binary classifiers´ performance confidence with the help of confusion matrix at first. Then, reconstruct the mass function based on the structure features of binary ECOC. Finally, the evidence for each class is got, which is determined not only by the outputs of binary classifiers, but by the classification performance for different patterns. Experimental results on UCI datasets with support vector machine(SVM) as the binary classifiers and the bias-variance analysis indicate that our approach can provide a better classification performance and promote the generalization ability.