Approximation sets of rough sets use existing knowledge granules to describe the uncertain concept approximately. However, the cost information contained in data have not been considered when constructing the approximation set of the uncertain concept. Therefore, the necessity of considering cost information is analyzed when constructing the approximation sets of rough sets firstly. Then an approximation set model of rough sets of misclassification cost is constructed from the perspective of cost-sensitive, and some properties of this model are discussed in detail. In order to find a suitable granularity space to describe the uncertain concept approximately that can minimize the sum of misclassification cost and test cost as far as possible in multi-granulation spaces, the contribute rate of attribute cost is defined, and the cost-sensitive granularity optimization algorithm is proposed. The experimental results show that the proposed algorithm can be used in existing cost cognition scene. And a reasonable hierarchy granulation space and the approximation set of the uncertain concept can be obtained under the given cost scene.