Learning from users: a data-driven method of QoE evaluation for Internet video

Improving quality of experience (QoE) is increasingly significant for Internet video content providers. The essential issue is how to evaluate QoE under the complex circumstance of Internet video. Based on the massive user data extracted from a large scale Video-on-Demand (VoD) provider, we present a data-driven, comprehensive and extendible study on the problems of QoE evaluation. The main works of this paper include obtaining QoE-associated features via feature engineering and building an evaluation model on features of different aspects for Internet video QoE. Firstly, for feature engineering, we propose to introduce pattern features of user viewing behaviors that interact with user-perceived video quality. A new method of frequent time series pattern mining is proposed to find typical patterns. We correlate user experience with user-perceived quality features and user behavior pattern features, and consider the impact of confounding factors by applying them as context features into modeling. Secondly, interdependency among features is challenging for QoE evaluation modeling. And the high dimension of feature vector should be considered. To address these challenges, we develop an ensemble method to model the interactions between features and their intricate relationships to user experience. Experiments demonstrate that our approach could achieve sound results in comparison with other related works.