Standard Data Envelopment Analysis (DEA) models are deterministic, and previous studies have made numerous efforts to introduce statistical analysis into DEA, such as the bootstrapping algorithm and regression-based approaches. In this study, we consider probabilistic variations present in input-output vectors and propose a forest-based sampling procedure to handle the statistical properties of efficiencies across different orientations of the DEA model. To capture the probability distributions of the data, we classify observed decision-making units into several clusters and maximize the information gain of each cluster using a Gaussian-based entropy function. The proposed approach is illustrated using a data set used in previous studies.