A new hierarchical framework is preliminarily proposed for accurate classification in imprecise multi-class domains inherently characterized by rarity and noise. The key idea behind the devised framework is coupling the individual rules of an associative classifier with as many local probabilistic generative models. These are trained over the coverage of the associated rules, wherein it is likely that some globally rare cases/classes become less rare. The individual local probabilistic generative models are then employed into the classification process for accurately dealing with the corresponding forms of rarity. Two novel schemes for a tight integration between associative and probabilistic classification are introduced, wherein the class of an unlabeled case is decided by considering multiple class association rules as well as their relative score produced by the probabilistic classifier. An intensive evaluation shows that the proposed framework is in most cases superior in performance w.r.t. an established rule-based competitor.
