Posterior Probability Reconstruction for Multi-Class Support Vector Machines

Pairwise coupling is a widely used method in multi-class SVM and max wins voting (MWV) strategy can obtain a global classification by considering each partial answer of binary classifier as vote. But MWV strategy has an important drawback, due to the nonsense caused by those meaningless binary classifier. This paper presents a novel approach, which considers the pairwise SVM classification as a decision-making problem and involves posterior probability to solve it. The combination strategy of the probability output among these binary SVM-based classifiers in one-against-one (OVO) decomposition is given. The strategy also considers the different prior probabilities of each binary classifier, which is evaluated by one-against-all (OVA) decomposition. The comparison is done with four benchmark data sets on UCI database and the performance of the proposed reconstruction strategy is validated with experimental results.