A Multiclass SVM Method via Probabilistic Error-Correcting Output Codes

Error-correcting output code (ECOC) is an effective approach to solve the problem of multiclass SVM. In this paper, a probabilistic approach that is based on ECOC is proposed. In the training stage, a coding scheme is predefined, and a special model is trained by samples. In the classification stage, besides the labels from SVM as usual, posterior probabilities of labels are also calculated. They are used to compute probability estimates of categories. Rank the normalized scores of probabilities and choose the maximum as the object category. Evaluations on different text categorization collections show our approach can significantly improve the performance.