DIG: A Data-Driven Impact-Based Grouping Method for Video Rebuffering Optimization

Shengbin Meng, Chunyu Qiao, Junlin Li, Yue Wang, Zongming Guo

Abstract: Rebuffering is known to be the dominant metric that affects the user experience of video streaming applications. In this paper, we propose a data-driven impact-based grouping (DIG) method for video rebuffering optimization. By analyzing data of 74.5 million video sessions collected in a real video streaming system, several key features with most significant and temporally persistent impact on video rebuffering are identified. Based on the values of these features, similar video sessions are grouped together. Within each group, we forecast future rebuffering events via a simple and efficient model, exploiting the insight that all video sessions in the same group face a similar risk of rebuffering. If rebuffering is predicted to happen in a coming session, we try to avoid it by selecting a better content distribution network (CDN) for this video. Experimental results show that our method can successfully predict $$46\%$$ of the rebuffering sessions, and reduce the average rebuffering rate by $$18.4\%$$ .