http://blog.csdn.net/luo123n/article/details/48878759
https://en.wikipedia.org/wiki/Hinge_loss Note that y should be the "raw" output of the classifier's decision function, not the predicted class label.