Traditional researches on the classification problem concern that a complete dataset is given as a training set without missing. However, incomplete data usually exist in real-world applications. In this paper, to handle incomplete numerical data in the classification problem, we propose a new approach based on fuzzy entropy. The proposed approach of handling incomplete data uses the technique of granular processing of fuzzy similarity measure to fill missing values of attributes. The experiments were made and the results were compared with the method of AMSC (attribute mean with same concept) through a few famous classification models to evaluate the performance of the proposed handling method.