The number of users streaming mobile video over the Internet has increased at an unprecedented rate throughout the past decade. Because video streaming is a mainstay of mobile computing, it is important that the best possible experience is delivered to users. Many mathematical algorithms have been proposed to improve quality in video-streaming-related domains. Typically, the parameters of these algorithms are established based on machine-centric indicators of video quality. Although researchers have attempted to connect these indicators with true user-perceived quality, in practice, there is often a disconnect. This project aims to improve directly-measurable indicators of user satisfaction in mobile video streaming by taking into account both individual user preferences as well as a user's tolerance for less than perfect quality in a specific video.
The proposed research will improve user-specific indicators by connecting problems in video streaming with problems in multi-task learning and collaborative filtering. These machine learning strategies are especially effective for prediction when large amounts of data are available. This effectiveness on large-scale learning fits well into this project's proposed context of improving user-facing indications of satisfaction. These indications include video abandonment, video session times, and collected navigation commands. Unlike user-surveys, these metrics can be collected at large scales through automated tooling in the video player. This research will investigate strategies that combine these large scale measurements with predictions from machine learning approaches toward selecting algorithm parameters that produce the most improvement in user-facing quality. This project will explore such parameter selection strategies in the context of improving user-perceived dynamic adaptive streaming quality. It will also explore such strategies to maintain a fixed level of user-facing quality while reducing mobile display power consumption via backlight scaling. Demonstrations of the approaches produced by this project will be featured in courses at the PI's institution and will be used to draw undergraduate interest toward computer science research.
