Improving QoE prediction in mobile video through machine learning

Despite the massive adoption of HTTP adaptive streaming technology, buffering is still the most harmful event for QoE in video streaming. Previous studies have shown that buffering is not only detrimental for the overall user experience, but is also highly correlated to viewer engagement. The occurrence of buffering is particularly critical in cellular networks and mobile video deployments, as network conditions are less stable and network resources more limited. In this context, monitoring and properly predicting the QoE of video streaming services becomes paramount to cellular network operators, who need to offer high quality levels to reduce the risks of customers churning for quality dissatisfaction. In this paper, we present a novel approach to multi-dimensional QoE prediction in mobile video using machine learning models. Contrary to previous models for QoE prediction in video streaming, which are generally uni- or low-dimensional and model the impact of single video descriptors independently, we use a high-dimensional input space to model the impact of buffering and initial delay on QoE. We train and test the proposed models on a publicly available mobile video dataset, generated from subjective QoE tests with real viewers. Besides improving prediction performance, the proposed models show that there is a clear influence of other buffering pattern descriptors generally neglected in previous models - in particular those linked to the occurrence of the last stalling event, shedding light on new KPIs to monitor for better QoE assessment in video streaming.