RT-DriftSelect: A Real-Time Dynamic Feature Selection Framework for Concept Drift Adaptation in SDN-IoT Environments

Abstract

The dynamic nature of Software-Defined Networking (SDN)-based Internet of Things (IoT) environments poses a major challenge for maintaining model accuracy under evolving data distributions, known as concept drift. This paper introduces RT-DriftSelect, a real-time adaptive framework designed to address this issue through dynamic feature selection and incremental learning. The proposed method continuously monitors feature statistics and computes drift magnitudes between consecutive data windows to identify relevant and evolving attributes. Features with significant drift are retained for model updates, while irrelevant ones are discarded, thereby enhancing computational efficiency. The retained features are incrementally trained using an AdaBoost–Hoeffding Tree ensemble, allowing continuous adaptation without retraining from scratch. Experiments conducted on the SDN-IoT dataset demonstrate that RT-DriftSelect achieves a high cumulative accuracy of 99.85%, while dynamically reducing the feature space by up to 93.8%, outperforming static baselines. The proposed system ensures both robustness and efficiency, making it suitable for real-time and resource-constrained IoT scenarios.