Décider à partir d’exemples de test partiellement connus

We consider classifying partially uncovered test instances, by uncovering (some of) their missing values via successive requests. The request budget is limited : the problem is then to concentrate on the most informative features. We study a strategy where the feature distribution is updated according to the uncovered values. After a few experiments, we briefly discuss the case of imprecise answers, and point out the difficulties which then arise.