Hybrid Machine Learning Classification and Inference of Stalling Events in Mobile Videos
Sirine Taleb, Nadine Abbas · 2022
Mobile video streaming accounts for a considerable percentage of network traffic. However, the fluctuations in the available bandwidth cause video stalling events which negatively affect the user’s quality of experience. The recognition of stalling events is of great importance for solving this issue. Predicting stalling events helps in predicting the user’s experience as well as aids in finding solutions to mitigate the existing issues. Despite the existence of some research that attempts to predict video stalling, no one framework proposes a hybrid approach that predicts this issue at several levels. In this paper, we propose a novel framework to classify video stalling into three levels while considering several supervised learning methods. First, logistic regression is used to detect the existence of a stalling event using a binary approach. Then, a supervised Random Forest model is trained to classify using multi-class classification the level of stalling events existing in a specific video. Moreover, beyond the acquisition of discrete video stalling level, a novel continuous quantitative method is proposed based on artificial neural networks to predict the number of video stalling events. For our experiments, we rely on a recently published dataset that collects 1,081 hours of time-synchronous video measurements at the network, transport, and application layer. Finally, the classification and regression results demonstrate the feasibility and accuracy of up to 92% of the proposed hybrid learning methods.