<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText">This is an announcement of Yihua Cheng's MS Presentation<br>
===============================================<br>
Candidate: Yihua Cheng<br>
<br>
Date: Thursday, May 11, 2023<br>
<br>
Time: 10 am CST<br>
<br>
Remote Location: <a href="https://uchicago.zoom.us/j/3150893650?pwd=RFgxNEE2MUQ0QzlsVXF3Ym94bDQ2Zz09">
https://uchicago.zoom.us/j/3150893650?pwd=RFgxNEE2MUQ0QzlsVXF3Ym94bDQ2Zz09</a><br>
<br>
Location: JCL 390<br>
<br>
M.S. Paper Title: Online Profiling and Adaptation of Quality Sensitivity in Internet Video<br>
<br>
Abstract: Video streaming systems separate two processes: (1) online video streaming (which optimizes quality metrics, such as higher bitrates, and fewer stalls), and (2) offline modeling of quality sensitivity (i.e., how the quality metrics affect average
user experience). As bandwidth scarcity and resource contention worsen, it is pressingly needed to better allocate resources by finer-grained modeling of how quality sensitivity varies during each video. However, per-video quality-sensitivity modeling has
been impractical, especially for live videos, as traditional offline user studies can be too slow for a new video before viewers watch it.<br>
We explore an alternative architecture, where quality sensitivity is modeled online by analyzing user actions in real<br>
video sessions as they stream the same video. The challenge is how to model quality sensitivity reliably and apply it in near-realtime to improve concurrent and future video sessions. We address the challenge in the context of SensitiFlow, a controller that
orchestrates adaptive-bitrate (ABR) logic of video sessions to optimize the common user-satisfaction metric of user engagement (view time per session). SensitiFlow creates an online control loop that (i) gradually profiles quality sensitivity per video segment
as more user experience-related feedback (e.g., exit or skip) is received from video sessions, and (ii) optimizes the ABR decisions of the video sessions to jointly improve their user engagement and generate more feedback. SensitiFlow’s control loop is fast
enough to profile quality sensitivity online and optimize bitrate decisions under common viewer arrival patterns of live events (e.g., live sports and TV shows). Using the real traces collected from 7.6M video sessions, we show that compared to a state-of-the-art
(baseline) ABR logic agnostic to the variation of quality sensitivity within a video, SensitiFlow (without using more bandwidth) can realize 80% of the improvement in engagement that would have been obtained by a hypothetical “oracle” system having the knowledge
of quality sensitivity in advance. Our user study also confirms that SensitiFlow can improve the mean opinion score (MOS) by 40% over the baseline ABR logic, suggesting that SensitiFlow’s online profiling of quality sensitivity is effective.<br>
<br>
Advisors: Junchen Jiang<br>
<br>
Committee Members: Junchen Jiang, Nick Feamster, Vyas Sekar<br>
<br>
</div>
</span></font></div>
<div class="BodyFragment"><font size="2"><span style="font-size:11pt;">
<div class="PlainText"><br>
<br>
</div>
</span></font></div>
</body>
</html>