“…Although they operate on the same class of functions and their learning problems all extend in a straightforward way the one of bi-class SVMs (precisely the 1-norm and the 2-norm ones), they exhibit distinct properties. In recent years, several comparative studies between M-SVMs and decomposition methods have been published (see for instance Guermeur, 2002;Hsu and Lin, 2002). In short, they establish that in practice, no model is uniformly superior or inferior to the others with respect to the standard criteria: prediction accuracy, sparsity, computational complexity, etc.…”