Considering multiple criteria classification problems,dominance relations and equivalence relations can be respectively introduced to condition attributes and decision attributes to describe different types of data. Based on the dominance-equivalence relations,a novel attribute reduction method based on sample pair selection is developed to deal with this kind of information systems. Instead of calculating the whole discernibility matrix,the proposed method only store the useful attributes for attribute reduction by selecting the discerned sample pairs,and therefore it can significantly improve the time costin attribute reduction. In addition,we propose an approximate reduction algorithm in order to deal with comparative large-scale information systems. This algorithm add attributes based on attribute importance and it's time saving. Finally,the experimental results on UCI data sets demonstrate the feasibility and effectiveness of the proposed algorithms.