Energy-Efficient Feature Selection for IoT Data in Cloud-Based Machine Learning Models
Azar, Kamali (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052114020
https://urn.fi/URN:NBN:fi:amk-2025052114020
Tiivistelmä
The rapid proliferation of Internet of Things (IoT) devices has amplified energy sustainability challenges in cloud-based systems, particularly when processing high-dimensional datasets for machine learning.
Conventional feature selection approaches often neglect energy constraints, increasing operational costs in resource-limited IoT environments. This study introduces a two-stage framework combining statistical relevance filtering and swarm-based optimization to address this gap. The first phase employs correlation analysis to discard nonpredictive features, minimizing data transmission overhead. The second phase integrates Particle Swarm Optimization (PSO) with energy-aware cost functions to dynamically refine feature subsets, balancing computational efficiency and model accuracy. Evaluated on the TON-IoT dataset, the framework achieves a 20% reduction in energy expenditure and 25% faster processing speeds compared to baseline methods, while sustaining 92.5% classification accuracy. Key innovations include GPU-accelerated PSO for real-time adaptability, edge-cloud preprocessing to reduce transmission loads, and empirical validation of energy-accuracy trade-offs in heterogeneous IoT scenarios. Practical implementation strategies are proposed for scalable deployment in smart cities and industrial IoT networks, emphasizing reduced carbon footprints. This work advances sustainable IoT ecosystems by harmonizing machine learning efficacy with energy conservation, providing foundational insights for future research in adaptive thresholding and federated learning architectures.
Conventional feature selection approaches often neglect energy constraints, increasing operational costs in resource-limited IoT environments. This study introduces a two-stage framework combining statistical relevance filtering and swarm-based optimization to address this gap. The first phase employs correlation analysis to discard nonpredictive features, minimizing data transmission overhead. The second phase integrates Particle Swarm Optimization (PSO) with energy-aware cost functions to dynamically refine feature subsets, balancing computational efficiency and model accuracy. Evaluated on the TON-IoT dataset, the framework achieves a 20% reduction in energy expenditure and 25% faster processing speeds compared to baseline methods, while sustaining 92.5% classification accuracy. Key innovations include GPU-accelerated PSO for real-time adaptability, edge-cloud preprocessing to reduce transmission loads, and empirical validation of energy-accuracy trade-offs in heterogeneous IoT scenarios. Practical implementation strategies are proposed for scalable deployment in smart cities and industrial IoT networks, emphasizing reduced carbon footprints. This work advances sustainable IoT ecosystems by harmonizing machine learning efficacy with energy conservation, providing foundational insights for future research in adaptive thresholding and federated learning architectures.