Video QoE Estimation using Network Measurement Data

More than even before, last-mile Internet Service Providers (ISPs) need to efficiently provision and manage their networks to meet the growing demand for Internet video (expected to be 82% of the global IP traffic in 2022). This network optimization requires ISPs to have an in-depth understanding of end-user video Quality of Experience (QoE). Understanding video QoE, however, is challenging for ISPs as they generally do not have access to applications at end user devices to observe key objective metrics impacting QoE. Instead, they have to rely on measurement of network traffic to estimate objective QoE metrics and use it for troubleshooting QoE issues. However, this can be challenging for HTTP-based Adaptive Streaming (HAS) video, the de facto standard for streaming over the Internet, because of the complex relationship between the network observable metrics and the video QoE metrics. This largely results from its robustness to short-term variations in the underlying network conditions due to the use of the video buffer and bitrate adaptation. In this thesis, we develop approaches that use network measurement to infer video QoE. In developing inference approaches, we provide a toolbox of techniques suitable for a diversity of streaming contexts as well as different types of network measurement data. We first develop two approaches for QoE estimation that model video sessions based on the network traffic dynamics of the HAS protocol under two different streaming contexts. Our first approach, MIMIC, estimates unencrypted video QoE using HTTP logs. We do a large-scale validation of MIMIC using ground truth QoE metrics from a popular video streaming service. We also deploy MIMIC in a real-world cellular network and demonstrate some preliminary use cases of QoE estimation for ISPs. Our second approach is called eMIMIC that estimates QoE metrics for encrypted video using packet-level traces. We evaluate eMIMIC using an automated experimental framework under realistic network conditions and show that it outperforms state-of-the-art QoE estimation approaches. Finally, we develop an approach to address the scalability challenges of QoE inference. We leverage machine learning to infer QoE from coarse-granular but light-weight network data in the form of Transport Layer Security (TLS) transactions. We analyze the scalability and accuracy trade-off in using such data for inference. Our evaluation shows that that the TLS transaction data can be used for detecting video performance issues with a reasonable accuracy and significantly lower computation overhead as compared to packet-level traces.Ph.D