In this paper, we place ourselves in a machine learning context and we tackle the problem of the ranking of attributes through bipolar information. From a classical training set, bipolar sets (Atanassov intuitionistic fuzzy sets or interval-valued fuzzy sets) are constructed and could thus be used with bipolar information measures, such as entropies, in order to produce a novel approach to rank attributes. With such an approach, new means to highlight the lack of knowledge associated with the distribution of attribute values are offered.