Decision-making of label prediction has a significant influence on classification performances, especially in case of high uncertainty. In classification tasks, classifiers are trained to predict the labels of new instances, implying a decision making process. In this paper, various decision criteria in evidential classification are analysed comprehensively in the belief function framework. Since it will be more cautious to make set-valued predictions with limited information, two approaches that based on a complete preorder among partial assignments and on a partial preorder among precise assignments are emphasized theoretically. Decision-making and performance evaluation in uncertain situations both generate the demand of an extended utility matrix. To define the utility of set-valued prediction, we propose two methods based on Ordered Weighted Average (OWA) operators and discounted utility mapping. To deal with uncertainty, the belief function framework is selected among various mathematical models in this paper, as it has the ability to express a variety of information availabilities, ranging from full information to complete ignorance.