Avancées récentes dans le domaine de l'apprentissage d'ordonnancements

In a wide variety of applications, where the data X ∋ χ that must be processed characterize instances to which binary labels Y ∋ {-1, +1} are randomly assigned, the goal of statistical learning does not reduce to find the likeliest label for a given instance but consists in ranking all the instances x ∋ χ in the same order as the one induced by the probability a posteriori η (x) = P {Y = +1 | X = x}, ranking rules being evaluated through ROC analysis. In contrast to the majority of procedures used in practice, based on a preliminary estimation of the function η (x), the results described in this article propose an extension of the concept of decision tree to the ranking problem in order to optimize the ROC curve directly.