Analysis And Application Of User Behavior In Crowdsourced Live Video Streaming

Crowdsourced live video streaming,which attracts vast number of users by its rich viewer-broadcaster interaction mechanism,has flourished and expanded over the past few years.The analysis of live video streaming platform has become a research hotspot in the field of streaming media services.In crowdsourced live video streaming plat-form,viewers are allowed to post comments to interact with the broadcasters,and buy virtual-gifts provided by the platform to reward the broadcasters.The existence of these interactive mechanisms makes the viewer not only a consumer of the content,but also a part of the content production.Understanding the interactions between viewers and broadcasters are essential for people to comprehend the production and consumption of the crowdsourced live video content,and improve the services.In this dissertation,taking Douyu,which is one of the most popular crowdsourced live video streaming platform,as a case study,we deeply analyze the user behavior in crowdsourced live video streaming.We crawled more than four months of data,and built a suite of models to analyze the user behavior,including the viewers' and the broadcasters'.It mainly includes the following four aspects.First,in-depth analy-sis of viewer behavior.We analyze viewer behavior in danmu comments posting and virtual-gifts donating using powerlaw,and it is found that their distribution is heavy-tailed.We conduct clustering analysis on the high consumption viewer community,the result shows that high consumption viewer community can be obviously clustered into three categories.Second,in-depth analysis of broadcaster behavior.By analyzing the attractiveness and profitability of professional broadcasters,it is found that their rank-ing distribution can be described by Zipf distribution.From the point of view of game broadcaster and show broadcaster,we make a comparative analysis of broadcaster's profitability,and find that the profitability of the show broadcaster is generally higher than that of the game broadcaster.Third,highlight detection in live broadcasting.Un-like the existing methods of highlight detection,which mostly focus on audio or video data itself,we propose an automatic highlight detection method based on the time series of danmu comment quantity and virtual-gift value in the broadcasting.The result shows that the proposed method can identify highlights with high accuracy.Fourth,channel popularity analysis.We propose that in crowdsourced live video streaming,besides the traditional viewer popularity based on the number of real-time online viewers,the channel popularity includes the danmu popularity and gift popularity derived from the interaction between the viewer and the broadcaster.We fit the distribution of the three types of channel popularity,and the result shows that their distribution does not follow Zipf distribution,but follows the stretched exponential distribution.The measurement and analysis in this dissertation has important implications on the strategy optimization and service improvement of crowdsourced live video streaming service.