Classification by pairwise coupling of imprecise probabilities

In this paper, we are interested in making decisions by combining classifiers providing uncertain outputs, in the form of sets of probability distributions. More precisely, each classifier provides lower and upper bounds on the conditional probabilities of the associated classes. The classifiers are combined by computing the set of unconditional probability distributions compatible with these bounds, by solving linear optimization problems. When the classifier outputs are inconsistent, we propose a correcting step that restores this consistency. The experiments show the interest of our approach for solving multi-class classification problems, particularly when information is scarce (i.e., a limited number of classifiers is available). In this case, modeling the lack of information associated with classifier outputs gives good results even when they are poorly regularized or overfit the data.