<p>Attribute reduction has consistently been a significant area of application for rough set theory. In this paper, we put forward a novel multi-granularity variable precision fuzzy rough set (which includes three basic models: optimism, pessimism, and compromise) and present its properties. Then, we investigate the concepts of approximate quality, uncertainty degree and combined approximation quality-uncertainty degree by utilizing the novel multi-granularity variable precision fuzzy rough set and provide the corresponding reduction methods. Besides, we design the heuristic reduction algorithms in accordance with the significance of attributes. We offer examples of credit card applicants to show the algorithm flow. At last, we carry out comparative experiments on eight open source data sets. The results demonstrate that the proposed reduction methods can effectively eliminate redundant attributes, have acquired high classification accuracy on KNN, SVM, and CART classifiers, and possess better reduction performance than some existing models.</p>