Recent research efforts have addressed the challenge of estimating HTTP adaptive video streaming Quality of Experience (QoE) and Key Performance Indicators (KPIs) from a network provider perspective, commonly relying on machine learning models and the analysis of features extracted solely from encrypted network traffic. This challenge is further complicated in light of realistic end user playback-related interactions, such as video skipping, pausing, and seeking. Given that user interactions impact traffic characteristics, such scenarios need to be considered when training QoE/KPI estimation models. We train models on datasets with and without user interactions (focusing on YouTube as a case study), with the aim to investigate the impact of user interaction on classification accuracy. Results motivate the need to systematically include data corresponding to various interaction scenarios when training QoE/KPI classification models that would be applicable in real-world scenarios.
展开▼
