Short Video Characterization: Design, Implementation and Evaluation

<p>Short videos have recently emerged as a popular form of User-Generated Content (UGC) within modern social media. This content, typically less than a minute long, is predominately produced in vertical orientation on smartphones. While still fundamentally streaming, short video delivery on mobile devices is distinctly characterized by the aggressive usage of pre-loading. While still fundamentally streaming, short video delivery on mobile devices is distinctly characterized by the aggressive usage of pre-loading portions of the next videos for a user. This pre-loading helps to overcome the varying dynamics of wireless connectivity and also allows for a nearly instant fast swiping to the next video. Furthermore, the prevalent usage pattern of rapid content swiping as enabled by pre-loading results in significant wastage by virtue of streaming buffers or pre-loads never being watched. Given the heavy cost of video and limited bandwidth afforded to wireless devices, managing the tradeoffs between user Quality of Experience (bitrate, video stalls, video start-up delay) and data wastage/browsing seamlessness is complex and multi-faceted. It is this problem that forms the foundation for this dissertation understanding the current state of short video and providing solutions that optimize these tradeoffs associated with short video delivery. </p><p>To accomplish that, this research first delves deeply into the selection of mobile video bitrates, presenting a low-consumption solution from a bandwidth prediction perspective. Next, this research critically evaluates the pivotal factors and design choices of existing short video platform pre-loading mechanisms, while also considering the impact on user Quality of Experience (QoE), bitrate, and perceived video quality. Finally, this research introduces a Kettle Buffer Loading (KBL) dynamic pre-loading algorithm. KBL is designed to sense network conditions and employ an adaptive pre-loading strategy, aiming to enhance video QoE while minimizing data wastage. The dissertation continues by evaluating KBL and state-of-the-art algorithms for short video delivery, carefully exploring the benefits and pitfalls of prediction modeling and vendor network scenarios.</p>