Real-time mobile broadband quality of service prediction using AI-driven customer-centric approach

Abstract

Statistical methods employed in evaluating the quality of service (performance) of mobile broadband (MBB) networks face drawbacks relating to the accurate and reliable processing of the huge amounts of heterogenous real time traffic data generated from MBB networks. Since the traffic patterns experienced in MBB networks are largely complex, highly dynamic, and heterogenous in nature, statistical methods may not adjust adequately to the changing network conditions. This highlighted gap can be addressed by machine learning (ML), as it has been effectively used in the past to support the analysis and knowledge discovery of communication systems’
traffic data through the identification of intricate and hidden patterns. This paper presents the application of ML techniques to predict MBB Quality of Service (QoS) in real-time, using a custom-built MBB performance application referred to as MBPerf that collects five (5) network metrics (DNS lookup, download and upload speeds, latency, signal strength), location information, and device characteristics across diverse network conditions in the South West region of Nigeria. The QoS modeling task was carried out using an MBPerf pre-processed dataset. Three (3) classification algorithms, including Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were trained using the MBPerf QoS dataset and then evaluated in order to determine the most effective model based on certain evaluation metrics – accuracy, precision, F1-Score and recall. Following hyperparameter tuning to improve the model’s performance, the selected model was deployed in a realworld network environment to classify QoS into “Above Average,” “Average,” and “Below Average” categories. Mobile customers receive real-time notifications with actionable insights based on the predicted QoS class, empowering them to optimize their usage and troubleshoot issues. From the performance results obtained for
the 3 ML models trained with the MBPerf dataset, SVM (95%) and XGBoost (97%) significantly outperformed RF (59%) in terms of accuracy. However, the performance difference between the SVM and XGBoost models is not significant. Interestingly, the three models showed great capability to accurately make predictions on the three QoS categories (classes) as depicted by the ROC-AUC and mlogloss curves. Lastly, the feature importance plot shows that QoS is the collective effect of service performance and not a function of QoS metrics only that determines the degree of satisfaction of a user of the service. This Artificial Intelligence (AI) powered system promotes a more transparent and efficient MBB experience for all stakeholders in Nigeria’s fast evolving digital landscape.