Accurately estimating production frontiers remains challenging, especially with technological heterogeneity or multi-stage processes. We introduce an enhanced EATBoosting algorithm, a machine learning-based method that strengthens traditional Data Envelopment Analysis (DEA). Applied to Meta-DEA, our approach adopts a fairness-aware perspective by benchmarking group-specific frontiers against a unified meta-frontier, revealing technology gaps and disparities. In Two-Stage Network DEA, it promotes transparency by evaluating stage-wise efficiencies and uncovering internal bottlenecks. Using shape-constrained stochastic gradient boosting, our method improves generalisation, mitigates overfitting, and performs reliably with limited data. Simulations and empirical analyses show it outperforms traditional DEA, capturing nuanced efficiency patterns more accurately. By bridging machine learning and DEA, this work offers a robust, interpretable framework for fair and transparent performance assessment.