Analysis of Application-layer Data to Estimate the QoE of WebRTC-based Audiovisual Conversations.

Successful deployment of Web-based Real-Time Communication (WebRTC) applications needs appropriate Quality of Experience (QoE)-aware service management to assure acceptability from the user's perspective. To this aim, monitoring of application-level data was found to provide relevant insights to estimate the user's QoE. In this paper, we investigate the relationship between WebRTC session parameters (collected with the webrtc-internals tool) and the users' QoE (in terms of the Mean Opinion Score (MOS)) through in-depth statistical analysis aimed at identifying the most suitable parameters for QoE estimation. In this regard, we based on statistical metrics, Pearson Correlation Coefficient (PCC), and Analysis of Variance (ANOVA). Then, we trained three machine learning regression algorithms (Regression tree, Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP)) using the identified parameters as the input data and the MOS as the output to be predicted. Experimental results show that the statistical analysis based on the PCC identified the optimal set of WebRTC session parameters for estimating the end user's QoE. With this optimal set of features, the MLP achieved the greatest QoE estimation performance in terms of R2 (0.852) and Root Mean Square Error (RMSE) (0.282), outperforming state-of-the-art results.