Machine Learning QoE Prediction for Video Streaming over HTTP

Abstract

Streaming video services have been growing rapidly in the past decade due to the wide adoption of more capable mobile devices such as smartphones and tablets, together with the deployment of higher capacity mobile networks and more efficient video compression and streaming techniques. Dynamic adaptive streaming over hypertext transfer protocol (DASH) can adapt bitrates to fluctuating network bandwidth. This makes measurement of perceived video quality a significant task. Therefore, there has been a strong demand for Quality of Experience (QoE) measurements and prediction models. The work in this thesis focuses on improving the performance of existing QoE prediction models that still have a few limitations. We propose a feature enhancement approach for QoE prediction using machine learning (ML) methods. Given the need for an accurate and reliable data set to do our analysis, train and evaluate our machine learning models, we utilize the Waterloo Streaming QoE Database III (SQoE-III) that combines the effects of video compression, initial buffering, and stalling events along with the subjective user responses to them. Additionally, the database contains the results of implementing various objective video quality assessment (VQA) models. First, we perform an extensive data and correlation analysis to study the effects of the different QoE key influence factors (KIF) which account for the video quality degradation. We further study the effect of quality switching that includes switching up and down, switch magnitude, ratio/time spent on the highest quality level/layer, in addition to the effect of stalling expressed as the number of stalls and their duration and initial buffering time. We analyze the effect of various objective video quality assessment (VQA) models on the QoE. The results of the analysis reveal the dependency between objective VQA models features, playback stalling, and quality switching features that affect the QoE. Towards effectively predicting user QoE and based on the results of our analysis, we identify/introduce a new set of enhanced features that combine objective VQA features, playback stalling features, and quality switching features, as input to various machine learning models to predict the values of QoE. Experimental results show that the proposed models are in close agreement with subjective opinions and significantly outperforms existing QoE prediction models and can also provide a highly effective and efficient future for QoE prediction in video streaming services.