Estimating accurate multi-class probabilities with support vector machines

In this paper, we propose a comparison of several post-processing methods for estimating multi-class probabilities with standard support vector machines. The different approaches have been tested on a real pattern recognition problem with a large number of training samples. The best results have been obtained by using a "one against air coupling strategy along with a softmax function optimized by minimizing the negative log-likelihood of the training data. Finally, the analysis of the error-reject tradeoff have shown that SVM allows to estimate probabilities more accurate than a classical MLP, which is indeed promising in the view of incorporated within pattern recognition system using probabilistic framework.