Concept Drift Under Harsh Constraints: A Review of Potential Strategies for IoT Systems
Ask Espensønn Øren, Nina Kuna Peuker, Akira-Miranda Adeyomi Adeniran-Lowe, Sarah Ruepp, Martin Nordal Petersen · IEEE Access · 2025
Concept drift, referring to the temporal evolution of data distributions, presents a significant challenge to the deployment of machine learning models in dynamic, real-world environments. This paper surveys recent advances in autonomous concept drift detection and adaptation methods, with a particular focus on their applicability to resource-constrained platforms such as IoT devices. This survey aspires to be a guideline for choosing a suitable architecture for any situation given the trade-off among detection accuracy, adaptability, and resource efficiency. Our analysis reveals that while statistical methods offer low overhead, they are often insufficient for complex drift scenarios; conversely, more accurate methods may exceed embedded system limitations.We conclude that modular, context-aware approaches that co-optimize both detection and adaptation under tight hardware constraints are essential for practical deployment in embedded learning systems.