Using probability estimation via outputs of SVM in ECOC

The Error-Correcting Output Codes (ECOC) and Support Vector Machines (SVMs) are widely used in classification. Traditional decoding process in ECOC is a hard decision. In this paper, an algorithm of probability estimation via outputs of SVM in ECOC is proposed. First, an appropriate coding matrix is constructed to make sure that pseudoinverse exist. Second, the original problem is transformed into multiple two-class SVMs according to the matrix. Finally, the probability approximated by value of SVM decision function and pseudoinverse of coding matrix are used to estimate the target class. Experimental evaluations on two datasets show that the proposed algorithm can improve the performance of traditional methods.