La perte par paires est un type de fonction de perte used primarily in apprentissage automatique models, particularly for tasks involving ranking, classification, and l'apprentissage métrique. Unlike traditional des fonctions de perte that evaluate the performance of a model based on individual predictions, pairwise loss focuses on comparing pairs of input samples. The goal is to ensure that the model correctly ranks or differentiates between these pairs based on their relative features.
In practice, Pairwise Loss works by selecting two samples at a time—typically one positive sample and one negative sample. The model’s predictions for these samples are then compared. The loss is computed based on whether the model correctly identifies which sample should be ranked higher or classified as more relevant. This approach is particularly useful in applications such as systèmes de recommandation, information retrieval, and face verification, where the relationship between items is more critical than their individual scores.
Les types courants de perte par paire incluent :
- Perte contrastive: Used to minimize the distance between similar pairs while maximizing the distance between dissimilar pairs.
- Perte de charnière : Often employed in machines à vecteurs de support, it penalizes predictions that do not meet a certain margin of separation between classes.
By focusing on pairs, this loss function can improve the model’s performance in scenarios where the order or relative comparison is more important than absolute predictions. This makes Pairwise Loss particularly valuable in situations where the data is inherently comparative, allowing for more nuanced learning and better generalization aux données non vues.