La pérdida de bisagra es una popular función de pérdida primarily utilizado en aprendizaje automático, especially within the context of maximum-margin classification tasks. It is particularly associated with Máquinas de Vectores de Soporte (SVMs) y otros algoritmos que buscan separar los puntos de datos con un hiperplano.
La función de pérdida de bisagra se define como:
Loss(y, f(x)) = max(0, 1 - y * f(x))
Here, y represents the true label of the data point (either +1 or -1), and f(x) is the predicted value from the model. The hinge loss calculates the error based on how far the predicted value is from the correct side of the frontera de decisión. If the prediction is correct and sufficiently far from the margin (i.e., the model confidently classifies the data point), the loss is 0. However, if the prediction falls within the margin or is incorrect, the hinge loss increases linearly.
Hinge loss has distinct advantages in SVMs, as it encourages the creation of a robust model that not only classifies data correctly but also maximizes the distance between the decision boundary and the nearest data points. This property of maximizing the margin helps in achieving better generalization en datos no vistos.
Aunque la pérdida de bisagra es efectiva para tareas de clasificación binaria, it can be extended to multi-class problems using techniques like one-vs-all or one-vs-one approaches. Nevertheless, one should be cautious when applying hinge loss in cases where the data is not linearly separable, as the model may struggle to find an optimal hyperplane.