The Decline of Synthetic Oversampling in Large-Scale Imbalanced Learning:A Post-SMOTE Empirical and Theoretical Study (2020–2025)

For over twenty years, SMOTE has been the standard default for addressing class imbalance. Yet a striking paradox has emerged: while researchers continue citing SMOTE extensively, practitioners have largely abandoned it in production systems. To understand this disconnect, we conducted a large-scale empirical and theoretical study of 821 papers published between 2020–2025, revealing a fundamental paradigm shift. Approximately 30% of new solutions now employ Generative AI (Diffusion Models), 30% rely on cost-sensitive loss functions, while the remainder explore hybrid approaches. We mathematically show why SMOTE fails at modern scales: its O ( N ² ) complexity exhausts memory on billion-sample datasets, its nearest-neighbor logic distorts high-dimensional manifolds, and its CPU-bound design is incompatible with GPU pipelines. Beyond theory, we identify the novel “SMOTE Paradox”—the critical divergence between academic citations and real-world deployment. This systematic study maps the emerging post-SMOTE landscape, syn- thesizes theoretical foundations across three competing paradigms, and provides decision rules for practitioners. Empirically, we validate these findings on real-world fraud data (N = 284,807, im- balance 578:1), confirming that cost-sensitive learning achieves parity with oversampling (+ 0.29% F1-score improvement) while eliminating preprocessing overhead. Our contributions include: (1) quantitative documentation of the paradigm shift through systematic analysis, (2) mathematical proofs of SMOTE’s failure modes, (3) the novel SMOTE Paradox framework, and (4) empirical validation on production-scale data. We conclude with actionable guidelines for practitioners and identify two critical open problems for future research in large-scale imbalanced learning.