A Comparison of Classical Machine Learning and Deep Learning Methods with a Proposed Transformer-based Framework for Network Traffic Prediction, Congestion Control and Intelligent Routing in Wireless Networks

The demand for wireless network traffic is increasing due to the high-bandwidth applications, including video streaming, cloud computing, IoT, and next-generation communication systems, makes conventional network management approach more and more insufficient. Representative Classical Machine Learning (ML) models (i.e., Random Forest, Support Vector Regression, and XGBoost) and Deep Learning (DL) designs (i.e., Long Short- Term Memory (LSTM), Gated Recurrent Units (GRU), and Convolutional Neural Networks (CNN)) have been shown to improve the traffic prediction and congestion detection performance; however, all have critical weaknesses in modelling long-range temporal dependencies, adapting to rapidly changing network states, and dealing with the highdimensional nonlinear dynamics of the real world network traffic in wireless network. This paper presents a comparative review of ML and DL models, identifying their limitations to three fundamental core problems: network traffic prediction, congestion control, and intelligent packet routing. Based on this, proposed a Transformer-based ML framework that utilizes a multi-head self-attention mechanism to capture global temporal and spatial dependencies over traffic sequences, which directly addressing the limitations of the existing works. The proposed framework integrates a data collection module, a transformer model core, and an SDN-enabled adaptive feedback loop to enable proactive, context-aware network management. Comparative analysis show that the Transformer-based approach performs better than classical ML and DL baselines in prediction accuracy, routing efficiency and congestion awareness, serving as a scalable and intelligent basis for future wireless network control.