User dynamics and sensitivity analysis of meta-data on the YouTube social network

点击翻译按钮获取中文摘要

There is significant motivation to determine what descriptors of a YouTube video or channels maximize the the number of views. The goal of this research project is to apply the principle of revealed preferences together with statistical analysis and machine learning techniques to quantify how metadata (titles, descriptions, keywords, and thumbnails) associated with YouTube videos affects the static and dynamic response of YouTube users. A major contribution of this research is a non-parametric feasibility test to detect if YouTube users are utility maximizers in their response to meta-data. The feasibility test only requires the observable meta-data and responses of users. If uses are utility maximizer then their associated utility functions, which encode their preferences, can be constructed without prior assumptions on the utility. A non-parametric learning algorithm can then be used to predict the response of users to variations in meta-level data. Having estimated the utility function of YouTube users, we then use statistical and machine learning techniques to study how meta-level data effects the popularity dynamics of YouTube videos. We apply the feasibility test and machine learning techniques to a decade of YouTube data collected from BroadbandTV. The results provide valuable insight into which meta-level data to adjust to increase the popularity of videos.