Abstract: To explore the influence of important variables selected by different variable selection methods on the efficiency score of the DEA model, a variable importance method based on Q² is proposed, which does not need to assume that there is a linear relationship between variables, and the ordered homology pursuit LASSO is introduced, which does not ignore information variables with high correlation. The variables selected by the two methods are applied to the data envelopment analysis model (DEA) to explore the influence of important variables on efficiency. The results of numerical simulation and real data show that the variable importance method based on Q² outperforms other machine learning-based methods in most cases, and plays a significant role in improving the individual efficiency. However, in the scenario of a small sample size and a low variable correlation coefficient, the ordered homology pursuit LASSO is better. Therefore, choosing an appropriate variable selection method has an important impact on the efficiency of DEA models.