Impact of User Playback Interactions on In-Network Estimation of Video Streaming Performance

Abstract

With significant growth in video streaming services, coupled with widespread use of traffic encryption, network operators are faced with the challenge of monitoring Key Performance Indicators (KPI) needed to detect quality impairments and drive Quality of Experience (QoE) management mechanisms. Though literature suggests that QoE/KPIs can be inferred from encrypted network traffic using machine learning (ML) methods, most studies published thus far fail to account for frequent viewer interactions, such as pauses, seeking forward/backward, or video abandonment. Such playback-related interactions inherently impact traffic patterns used as input for ML–based KPI estimation models. In this paper, we investigate to what extent network operators can monitor application-layer KPIs considering realistic user interactions, focusing on the use-case of a popular streaming platform with videos streamed to mobile devices. We first investigate the impact of user interactions (pause, seek forward, and abandonment) on the performance of both session-based and real-time KPI classification models trained on datasets that do not contain interactions. Secondly, we systematically evaluate the performance of KPI estimation models trained on datasets including specific sets of interactions to determine which types of interactions need to be included in the model training procedure in order to be applicable for realistic streaming sessions.