This is a Matlab implementation of the fancy idea by Polson & Scott that reformulates the traditional binary linear SVM problem into a MAP (Maximum a Posteriori) estimation in a probabilistic generative model, and by use of the technique of data augmentation, makes it possible to do very easy and fast Gibbs sampling for the solution. You may find the paper here: "Data Augmentation for Support Vector Machines" by Nicholas G. Polson and Steven L. Scotty, published in Bayesian Analysis (2011) http://ba.stat.cmu.edu/journal/2011/vol06/issue01/polson.pdf Specifically, I implemented the basic EM algorithm and the MCMC algorithm, as well as the case under spike-and-slab prior. NEW: We extend the algorithm to the Crammer & Singer multi-class SVM! Check csmultisvm.m! The code is already highly optimized for Matlab (not further accelerated by MEX though).