Learning Contextual Discounting and Contextual Reinforcement from Labelled Data

This paper addresses the problems of learning from labelled data contextual discounting and contextual reinforcement, two correction schemes recently introduced in belief function theory. It shows that given a particular error criterion based on the plausibility function, for each of these two contextual correction schemes, there exists an optimal set of contexts that ensures the minimization of the criterion and that finding this minimum amounts to solving a constrained least-squares problem with as many unknowns as the domain size of the variable of interest.