The set-valued information system allows the object to have attributes multi-valued to achieve a more comprehensive description of complex information. In traditional set-valued information systems, each attribute has only one scale. But in the specific application process, people often need to process and analyze the data at different scales. In this paper, the granularity transformation function of a multi-scale information system is introduced into the set-valued information system, and the theoretical framework of the multi-scale set-valued information system is established. The relationship of information grains and rough sets between different scales of the system is discussed. On this basis, a multi-scale set-valued decision information system model is established. The transfer properties of consistent between different scales of the model are discussed. Then, we discuss the optimal scale selection method for a consistent and inconsistent multi-scale set-valued decision information system. The improved granular computing model of multi-scale decision information system has certain value in theoretical analysis and practical application.