Deep Learning and Optimization Approaches in Prediction of IoT Traffic Using Gradient Boosting, Auto-Metric Graph Neural Network, and Lyapunov Optimization-Based Predictive Model: A Review

Abstract

The rapid proliferation of Internet of Things (IoT) devices has significantly increased network traffic complexity, requiring intelligent and adaptive prediction mechanisms for efficient network management. Traditional traffic prediction approaches often fail to capture nonlinear, dynamic, and spatiotemporal dependencies inherent in IoT environments. This paper presents a comprehensive review of advanced deep learning and optimization-based techniques for IoT traffic prediction, focusing on Gradient Boosting methods, Auto-Metric Graph Neural Networks (GNNs), and Lyapunov Optimization-based predictive models. Gradient Boosting techniques provide strong performance in handling structured data and improving prediction accuracy through ensemble learning. Meanwhile, Graph Neural Networks effectively capture spatial relationships and dependencies among network nodes, enabling enhanced traffic forecasting in distributed IoT architectures. Lyapunov optimization offers a robust mathematical framework for real-time decision-making and dynamic resource allocation, balancing latency, energy efficiency, and throughput. Recent studies demonstrate that integrating deep learning with optimization techniques significantly improves prediction accuracy, reduces network congestion, and enhances resource utilization. However, challenges such as scalability, data heterogeneity, privacy concerns, and computational overhead remain critical issues. This review systematically analyses existing literature, highlights comparative strengths and limitations of various approaches, and identifies future research directions for hybrid intelligent IoT traffic prediction systems. The findings suggest that combining machine learning, graph-based modelling, and optimization frameworks is a promising direction for next-generation IoT networks, particularly in 5G/6G and edge computing environments.