Multi-Objective Evaluation of Classification Algorithms for Internet of Things (IoT) Systems and Edge Computing

Abstract

—The rapid rise of connected systems and the Internet of Things (IoT) has led to the widespread deployment of machine learning techniques in distributed and resource-constrained environments. In this context, the selection of classification algorithms can no longer rely solely on predictive performance, but must also consider computational cost, latency, robustness, and deployment constraints. This paper proposes a multi-objective evaluation framework for selecting supervised classification algorithms in IoT-oriented connected systems. A comparative theoretical study of major classification algorithms is first conducted using a trade-off-based perspective. The framework is then validated experimentally on two complementary datasets: the Human Activity Recognition (HAR) dataset, representing sensor-based IoT applications, and the University of New South Wales Network-Based 15 (UNSW-NB15 ) dataset, representing network-based IoT environments. The evaluation incorporates both predictive and computational metrics, including inference time, in order to better reflect real deployment constraints. The results confirm that no single algorithm dominates all criteria and that model selection strongly depends on the target deployment scenario, whether in constrained IoT nodes, edge computing platforms, or cloud environments. The proposed framework provides a structured decision-support tool for selecting classification algorithms in connected systems, highlighting the importance of multi-objective trade-offs for real-world IoT deployments. 