Evolving Machine Learning in Non-Stationary Environments: A Unified Survey of Drift, Forgetting, and Adaptation

In an era defined by rapid data evolution, traditional machine learning (ML) models often struggle to adapt to dynamic and non-stationary environments. In this work, we present evolving machine learning (EML) as a unifying paradigm for adaptive learning under distributional change, enabling continuous updating and real-time adaptation to streaming data. While prior surveys have examined individual aspects of evolving learning, such as drift detection or continual learning, there remains a lack of an integrative analysis that connects its major challenges within a coherent perspective. This survey provides a comprehensive review of EML by analyzing four interrelated challenges: data drift, concept drift, catastrophic forgetting, and skewed learning. We examine 147 recent studies, categorizing state-of-the-art approaches across supervised, unsupervised, and semi-supervised settings. Beyond taxonomic organization, we synthesize these challenges through a unified analytical perspective, highlighting their structural interdependencies and the inherent trade-off between adaptability and stability in evolving systems. Furthermore, we review evaluation protocols, benchmark datasets, and real‑world applications, offering a comparative assessment of methodological strengths and limitations while identifying key research gaps and emerging opportunities for robust and scalable EML systems.