Margin

A measure of how decisively a classifier favors one prediction over competing alternatives, expressed as a score gap or distance from a decision boundary.

Connect

Decision boundary: A surface in feature space that separates regions assigned to different predictions by a classifier. A geometric margin measures an example’s distance from this separating surface.
Classification: The task of assigning inputs to one of a finite set of categories. A classification margin describes separation among the candidate category predictions.
Maximum-margin classifier: A classifier trained to choose a decision boundary with the largest separation from labeled examples. Maximizing the minimum margin gives a principled way to select among separating boundaries.
Classification confidence: A model’s stated degree of certainty in a classification prediction, often represented by a probability. A large score margin may indicate confidence, but calibration is needed to interpret it probabilistically.
Hinge loss: A loss function that penalizes predictions whose scores fail to exceed a target by a specified margin. It turns insufficient separation into a direct training penalty.
Class score: A numerical value a classifier assigns to a class for a given input. The gap between the winning score and a competitor defines a common margin.
Margin-based learning: A family of learning methods that train predictors by encouraging correct predictions to exceed alternatives by a margin. It extends margin objectives beyond the standard support vector machine.
Probability calibration: The agreement between predicted probabilities and observed outcome frequencies. Calibration evaluates probability reliability, whereas a margin measures separation between scores or boundaries.
Support vector machine: A supervised learning method that selects a separating boundary by maximizing the margin between classes. Its optimization objective makes margin the defining quantity.
Multiclass classification: Classification in which each input may belong to one of more than two classes. Its margin usually compares the top class score with the strongest alternative.