Gradient Boosting for classification.
GB builds an additive model in a
forward stage-wise fashion; it allows for the optimization of
arbitrary differentiable loss functions. In each stage ``n_classes_``
regression trees are fit on the negative gradient of the
binomial or multinomial deviance loss function. Binary classification
is a special case where only a single regression tree is induced.