A Performance-Based Hybrid RT–HT Model-Switching Mechanism in 5G Handover Optimization

Abstract

The problem of concept drift poses a critical challenge in dynamic 5G networks, causing static machine learning models to fail when data distributions change. While online learning techniques can adapt, but suffer from cold start problems. This paper presents a performance-based Hybrid Regression Tree (RT) - Hoeffding Tree (HT) technique for concept drift detection in 5G handover optimization. The technique dynamically selects between a pre-trained Regression Tree (RT) and an incrementally trained Hoeffding Tree (HT) using a mobility-domain indicator and a sliding window of recent prediction errors. The study conducted an extensive simulation using the 3GPP TR 38.901 urban-microcell channel model and compared the proposed hybrid technique against five baseline techniques: ML-SOHOT (Alraih et al., 2025), Hoeffding Tree, LIM2 (Karmakar et al., 2022), E-SARSA Reinforcement Learning (Junejo et al., 2025), and PPO (Voigt et al., 2024). The results show that the proposed Hybrid technique reduces handover failure by 82.2% under drift conditions, achieves 44% reduction in handover latency, and 66% reduction in ping-pong rate compared to MLSOHOT (RT). The technique maintains consistent handover interruption time (1.8 ms) and achieves the best balance across all the KPIs among the compared baseline studies, confirming its effectiveness for real-time 5G deployment.