Could you please explain about algorithm = ‘SAMME’? how does this differ from ‘SAMME.R’?
The reference look at is the following where all details are given: https://web.stanford.edu/~hastie/Papers/samme.pdf
However, the “R” version stands for “Real”. In SAMME algorithm, it is expected that the weak-learner will provide thresholded prediction (class labels) as given by model.predict
in scikit-learn. The SAMME.R is another algorithm that relies on getting a continuous estimate to belong to a class as the probability as given by model.predict_proba
in scikit-learn.
If you look at Algorithm 2 and 4, you will see that taking the probability allows to not have to compute the alphas
. If I recall well, SAMME.R is stated to converge faster.