Feature selection using variable precision neighborhood rough sets (VPNRS) has garnered considerable attention in data mining and knowledge discovery. Nevertheless, the positive region of VPNRS may not be strictly divided due to the introduction of variable parameters, which could reduce the credibility of feature significance. Meanwhile, the calculation of approximate space is also complex and expensive. Hence, how to improve the computation efficiency is also an investigated issue. As for these issues, we propose a variable precision composite measure and design a novel local method for the feature selection of decision data. Firstly, we introduce the variable precision neighborhood rough set model to process uncertain information from global and local viewpoints. Furthermore, the variable precision composite measure is defined to evaluate the model’s accuracy and further used to select the essential features. Finally, a local forward algorithm is provided for feature selection to improve computing efficiency. All experiments on twelve datasets show that the local method is efficient, and the feature selection algorithm based on variable precision composite measure performs well in classification performance. Our work will provide a convenient tool for feature selection methods with uncertainty measures.