Three decades of DEA-machine learning integration (1996–2025): A bibliometric analysis, science mapping, and state-of-the-art review

This study presents the first review that unifies a bibliometric analysis with an in-depth synthesis of methodological advances at the intersection of data envelopment analysis (DEA) and machine learning (ML), covering 3 decades of scholarly works from 1996 to 2025. Drawing on 500 peer-reviewed publications indexed in Scopus , the review maps the field’s intellectual evolution, thematic structure, and collaboration dynamics using the Bibliometrix package in R . The results reveal an exponential surge in research activity after 2015, driven largely by contributions from Asia , particularly China and Iran , and methodological leadership from Europe , with Spain emerging as the global epicenter of frontier-learning innovations. Science mapping further uncovers four dominant thematic clusters and highlights a decisive shift from early ANN and SVM-based hybrids to deep learning (DL), ensemble models, and recent attention-based architectures. A parallel state-of-the-art (SOTA) synthesis shows the rise of post-DEA learning models such as Efficiency Analysis Trees (EAT), Convexified EAT (CEAT), Data Envelopment Analysis-Based Machines (DEAM), Support Vector Frontier (SVF), Convexified SVF (CSVF), SVF-Splines, Multivariate Adaptive Regression Splines (MARS), Relaxed Additive SVF (RASVF), Relaxed Additive convex SVF (RAcSVF), Adaptive Constrained Enveloping Splines (ACES), and unsupervised DEA ( u DEA), signalling a conceptual transition toward ML-driven, interpretable frontier estimation. By integrating scientometric evidence with methodological analysis, this review offers a complete survey of the DEA-ML landscape to date and provides a foundational roadmap for future theoretical development and applied research in dynamic efficiency modelling.