A Survey of Methods and Architectures for Prediction of IoT Traffic Using Gradient Boosting, Auto-Metric Graph Neural Network, and Lyapunov Optimization-Based Predictive Model

Abstract

The exponential growth of the Internet of Things (IoT) has resulted in massive volumes of heterogeneous and dynamic traffic, necessitating accurate prediction models for efficient network management. Traditional statistical approaches are insufficient to capture the nonlinear, temporal, and spatial dependencies inherent in IoT traffic. Consequently, advanced machine learning, deep learning, and optimization-based models have been widely adopted. This survey explores recent methods and architectures for IoT traffic prediction, focusing on Gradient Boosting, Auto-Metric Graph Neural Networks (AM-GNN), and Lyapunov optimization-based predictive models. Gradient Boosting techniques provide high accuracy and scalability through ensemble learning, while Graph Neural Networks effectively model spatial relationships by representing IoT devices as graph structures. GNN-based models are particularly effective in capturing hidden dependencies in multivariate time series data. Furthermore, Lyapunov optimization provides a robust framework for ensuring system stability and dynamic resource allocation by transforming optimization problems into queue stability problems. Recent hybrid approaches combining GNNs with optimization techniques have demonstrated significant improvements in prediction accuracy, latency reduction, and energy efficiency. This survey systematically reviews recent literature (2020–2023), analyses key methodologies, and highlights emerging trends such as federated learning, reinforcement learning, and hybrid AI models. It also identifies major challenges including scalability, computational complexity, and data heterogeneity. The study concludes with future research directions aimed at developing efficient, scalable, and intelligent IoT traffic prediction systems.