The tradeoff between higher efficiency and wider stability of performance map is still one of the bottlenecks to hamper the further research and development of advanced multistage axial-flow compressor. The recent rapid growth of computational resources and artificial intelligence has enabled data mining as one of the most effective and potential ways to gain a deep insight into the complex correlations between aerodynamic performance and three-dimensional geometry parameters. In the open literatures, however, few research works have been found on using the data mining that is independent of design optimization to extract priori design guidelines for multistage axial-flow compressor mainly due to the lack of proper data mining method focused on the interpretation of metamodel with full use of limited time-consuming computational fluid dynamics dataset. To tackle this issue, a metamodel-interpreted data mining framework is developed in which extreme gradient boosting (XGBoost) metamodel combined with Shapley additive explanation (SHAP) model are employed to locally interpret the feature importance of each sample in the computational fluid dynamics dataset and then extract the design guidelines in terms of the most influential geometry parameters and their beneficial variation directions. The developed method is applied to data mining of design guidelines for efficiency and stability enhancement of a front 3.5-stage transonic axial-flow compressor in ship-board gas turbine usage. The results show that the aerodynamic performance of the investigated multistage compressor is most sensitive to three-dimensional geometry parameters related to blade lean, blade twist, and variable stators. Specially, the variable stators mainly affect the stall margin at part speed. The blade lean mainly influences the adiabatic efficiency at design speed as well as the stall margin at both speeds, while the blade twist mainly influences the aerodynamic performance at design speed. New designs followed by the design guidelines are obtained and critical performance indicators related to the goals of the data mining task are verified. The stall margin at part speed is widened to 5.87% with adjustment of blade lean and twist and further to 23.31% with additional adjustment of variable stators. The peak adiabatic efficiency at design speed is improved by 0.06% in spite of extremely limited potential for efficiency enhancement of the original design. The present work is of scientific significance as well as industrial application value in the three-dimensional design optimization of advanced multistage axial-flow compressor at the affordable computational cost.