Towards conservative inference in credal networks using belief functions: the case of credal chains