Characterizing viewing engagement patterns in DASH

Dynamic adaptive streaming over HTTP (DASH) has seen a rapid increase in recent years. While opening a door for ensuring smooth video playback with bitrate adaptation, especially for unstable networks, DASH also requires a set of new streaming strategies (e.g., bitrate adaptation). A key of designing these strategies is to understand the viewing engagement patterns in DASH, which are the correlation between streaming strategies and Quality of Experience (QoE) of users. To characterize viewing engagement patterns, we use large-scale traces from a DASH streaming provider, based on which we measure users´ viewing patterns in DASH sessions, and the contextual factors (e.g., streaming device) that affect the viewing engagement. Based on our measurement insights, we further design a viewing engagement prediction model, to let streaming providers predict how long viewers will remain in DASH sessions given the streaming contexts. The effectiveness of the prediction model is verified by the real traces collected.