Probabilistic Outputs for Multiclass Support Vector Machines Based on M-Ary Classification

Abstract

To obtain a more overall and accuracy result, a method of probabilistic outputs for multiclass support vector machines based on M-ary classification (POSVM) is proposed in this paper. Comparing with the conventional hard decision multiclass support vector machines (SVMs), the proposed method can reserve almost all the information contained in the samples, which is more beneficial for the post-processing. Furthermore, the conventional one-against-one method requires K(K − 1)/2 SVMs to partition K classes, whereas the proposed method only requires ⌈log2 K⌉ SVMs for the same problems. In addition, for the cases where the classes are of different importance, a weighted SVM (WSVM) based on POSVM is proposed, which can obtain more reasonable results in the pattern recognition applications. The experimental results show that the proposed method provides a more overall and accuracy result than the conventional multiclass SVMs, and that it can be achieved easily by organizing a set of SVMs with a simple structure, especially for the problems with large number of classes. On the other hand, the results of the WSVM method are more logical and efficient than those of the POSVM, and the different classification rules can be designed by changing the weights according to the different cases. Therefore, the complexity of implementing weighted classifiers can be greatly reduced, and the design process is more flexible.