Enhancing QoE in software-defined networks via machine learning-based QoS-to-vMOS estimation

Abstract Efficient network traffic management plays a pivotal role in ensuring reliable data delivery and user satisfaction. In this study, we investigate the relationship between Quality of Service (QoS) parameters and the Video Mean Opinion Score (vMOS), a critical metric for assessing Quality of Experience (QoE). Using a simulation-based dataset, ten machine learning algorithms—Linear Regression (LR), Ridge Regression (RR), K-Nearest Neighbors (KNN), Decision Tree (DT), Gradient Boosting (GB), LightGBM, XGBoost, Multi-Layer Perceptron (MLP), Support Vector Regressor (SVR), and Random Forest (RF)—were employed to estimate vMOS from QoS metrics. Among them, the LightGBM model achieved the highest R² score of 0.9997, while RR lagged with 0.8894. Although LightGBM delivered superior predictive accuracy, the XGBoost model demonstrated a more balanced performance, offering a comparable R² value of 0.9996 with a significantly lower inference time of 0.0081s. In classification tasks (Logistic Regression (LogR), KNN, DT, GB, LightGBM, XGBoost, MLP, Support Vector Classifier (SVC), and RF), excluding MLP and LogR, all evaluated models attained accuracy scores exceeding 0.98. Confusion matrix analyses and runtime evaluations revealed that GB and LightGBM models consistently provided robust and efficient results. Furthermore, the simulation results underscored that early activation of the proposed model led to improved Structural Similarity Index (SSIM) values, thereby enhancing the resulting MOS. These findings suggest that leveraging machine learning to map QoS metrics to user-perceived QoE can significantly improve service delivery in software-defined networks (SDN), offering both scalability and adaptability to dynamic traffic conditions.