With the prevalence of big data, traditional data envelopment analysis (DEA) faces the challenge of handling large-scale computations. The DEA computation time can increase significantly as the number of decision-making units grows. In this study, we propose a novel approach to accelerate DEA computations involving voluminous data. The proposed method uses random forest classification to predict and search for the best-practice DMUs within the large-scale observations. Since best-practice DMUs are always of a smaller quantity, and they can determine the efficiency scores of all the remaining DMUs, identifying best-practice DMUs first reduces the programming size and the consequent computation time of the DEA model. The proposed method is termed as the DEA-RF method, which combines DEA and machine learning methods to reduce computational cost. Next, we test the effectiveness of the proposed method using numerical cases involving large-scale data. After computing the DEA scores of the DMUs in both the observed and simulated samples, we find that the proposed DEA-RF method can decrease computation time significantly, while ensuring an acceptable level of accuracy. Additionally, the larger the sample size is, the more time the model can save. The proposed DEA-RF method proves to be an effective solution to the long computation time problem of DEA models under big data contexts.