Metrics hingeloss

HingeLoss

The only metric here that reads a decision function — not a label, and not a probability, but the signed distance from a boundary that a support vector machine or a linear model produces.

A sample costs nothing once it is on the right side by a margin of 1, and its cost rises linearly from there. That margin is the whole point: a prediction that is right but barely is still charged, where ZeroOneLoss counts it as free. This is the loss an SVM actually minimises, which is why it is the one to report when tuning one.

Only the sign of the decision matters

The label is compared against the decision's sign, so relabelling the two classes cannot move the number — -1/1, 0/1 and 7/3 all give the same loss on the same decisions, and a test asserts it. posLabel says which label is on the positive side; scikit-learn infers the two classes instead.

On a truth carrying only one class the two disagree. scikit-learn maps every label to -1 through a LabelBinarizer that has nothing to contrast, and returns a number computed against the wrong side — measured 1.65 where the margins say 0.35. Here posLabel is a parameter, so that input is ordinary and answers 0.35. The divergence is the reference's degenerate case, not a choice made here.

The multiclass form is a different margin

MultiClass takes one decision per class and charges on the true class's decision less the best of the others — Crammer and Singer's multiclass hinge, which is what the reference computes. A sample costs nothing once its own class wins by 1.

Members

Member	What it does
HingeLoss.Score	The binary case, from one decision per sample.
HingeLoss.MultiClass	The same over one decision per class.