Learning conditional probabilities from incomplete databases - An experimental comparison