Fits Platt scaling: a logistic regression of the class on logit(p), by Newton's method with a
backtracking line search, at most 100 iterations. The targets are Platt's smoothed ones,
(N₊ + 1) / (N₊ + 2) for a positive case and 1 / (N₋ + 2) for a negative one, so a separable set
still has a finite fit. It has converged when the gradient is below 1e-9.
Fits Platt scaling: a logistic regression of the class on logit(p), by Newton's method with a backtracking line search, at most 100 iterations. The targets are Platt's smoothed ones, (N₊ + 1) / (N₊ + 2) for a positive case and 1 / (N₋ + 2) for a negative one, so a separable set still has a finite fit. It has converged when the gradient is below 1e-9.