In formal concept analysis based applications, controlling the structure of concept lattice is of vital importance, especially for big data, and is achieved via clarifying the granularity of attributes. Existing approaches for solving this issue are within the framework of classical formal concept analysis, which focuses on positive attributes. However, experiments have demonstrated that both positive and negative attributes exert comparable influence on knowledge discovery. Thus, it is essential to explore the granularity of attributes in positive and negative perspectives altogether. As a solution, we investigate this problem within the framework of three-way concept analysis. Specifically, we present zoom-in and zoom-out algorithms to obtain more particular and abstract three-way concepts, separately. Furthermore, we provide illustrative examples to show the practical significance of this study.