Efficient maximum projection of database-induced multivariate possibility distributions

Current research in the domain of inference networks, probabilistic as well as possibilistic, focuses on learning such networks from data. Learning inference networks consists in finding a decomposition of a multivariate probability or possibility distribution that is induced by a database of sample cases. An operation to be carried out several times during the execution of common learning algorithms is the computation of the projection of the database-induced probability or possibility distribution to a subset of the database attributes. This operation is trivial for the probabilistic case, but turns out to be a problem for the possibilistic one, since ad hoc approaches lead to wrong results or are very inefficient. In this paper we suggest an efficient method to compute maximum projections of database-induced possibility distributions, making real world possibilistic network learning feasible in the first place.