Enhancing 5G and 6G networks through a dynamic dual-stage machine learning heuristic framework for selecting UEs as UE-VBSs

Adapting mobile networks to the diverse and evolving demands of 5G and forthcoming 6G technologies requires flexible, efficient, and dynamic strategies—especially in ultra-dense environments and infrastructure-limited areas. This paper proposes a robust two-stage Machine Learning (ML) heuristic framework to dynamically select a group of User Equipment (UEs) to act as Virtual Base Stations (UE-VBSs) for network augmentation. In the first stage, Self-Organizing Maps (SOM) are employed to cluster UEs based on their spatial characteristics while preserving topological relationships, achieving a silhouette score of 0.64—a 30% improvement over conventional methods such as $K$-Means (0.46) and Mean-Shift (0.43). In the second stage, a Random Forest classifier enhanced via the Synthetic Minority Over-sampling Technique (SMOTE) attains an average accuracy of 97% and an F1-Score of 0.88 in identifying eligible devices to become UE-VBSs, outperforming recent frameworks that typically report accuracies ranging between 85% and 92%.

Comparative evaluation results demonstrate that our two-stage ML heuristic framework not only improves clustering accuracy and UE-VBS classification but also consistently outperforms state-of-the-art clustering methods in terms of network sum rate, power consumption, and scalability. Specifically, across all device densities (i.e., 200, 400, 600, 800, and 1000 UEs), our approach achieves the highest sum rate—peaking at nearly 1.8 billion bps (or 1.8 Gbps) at 1000 UEs—thus surpassing methods such as Affinity Propagation and Grid-based Clustering. Furthermore, by intelligently selecting UE-VBSs, the framework significantly reduces power consumption by effectively minimizing redundant transmissions and interference, making it an energy-efficient solution for large-scale 5G networks. Although the complexity of SOM clustering and Random Forest classification introduces higher computational overhead, the resulting improvements in throughput, energy efficiency, and scalability justify this cost, making it a robust and practical solution for real-world deployments. Validated on both synthetic and real-world datasets, our findings underscore the efficacy, scalability, and high impact of employing robust unsupervised and ensemble learning techniques for dynamic network optimization in next-generation architectures, delivering up to a five-fold increase in network sum rate under high-density conditions compared to state-of-the-art approaches like grid-assisted clustering and affinity propagation.