Machine Learning Analysis of Multi-Radio Access Technology Selection in 5G-NSA Network

The quantum increase in the number of mobile subscribers which resulted into an exponential growth of traffic across the mobile networks called for exploitation of new spectrum bands of 5G networks whose deployment still rely on support from underlying long term evolution 4G (LTE) networks in both stand-alone (SA) and non-stand-alone (NSA) architectures, this poses challenges on the choice of Radio Access Technology (RAT selection) between 4G (LTE) and 5G-NR networks to these ever increasing mobile users with respect to their geographical location, mobility and bandwidth requirement for efficient resource allocation, hence this study investigates joint user requirements and network constraints for appropriate RAT selection between 4G (LTE) and 5G-NR by recording live 5G-NSA radio measurements at 1.5 meter intervals over distance of 300 meters between a pedestrian user and the base-station, the problem (RAT selection) is formulated as a classification process, hence measured input parameters are analysed with standard supervised classification machine learning algorithms: Decision Tree (DT), Extra Tree (XTREE), Random Forest (RF), Gradient Boosting (GB), and eXtreme Gradient Boosting (XGBoost) to select an appropriate RAT. Evaluation of results show measure of accuracy of algorithm: DT at 91.82%, RF at 87.64%, XTREE at 86.75%, GB at 91.86%, and XGBoost at 93.86%, which showed XGBoost with highest accuracy value.Furthermore cross-validation metric is also used to test the efficacy of algorithms on future data which also revealed that DT, RF, XTREE, GB and XGBoost have the following cross- validation score: 0.909, 0.912, 0.918, 0.918, 0.929 respectively where XGBoost again shows highest 92.9% validation score, hence XGBoost recommended as classification supervised machine learning model for RAT selection to achieve effective and efficient resource allocation