TEXT

Abstract

Predicting popularity of YouTube videos is a relevant problem for advertisers and content creators. Previous literature discusses predicting video popularity using a variety of features, including early number of views, video contents, and subscriber networks as well as external social media data on user comments and shares. However, there is limited research on video popularity resulting from the network of related YouTube videos, specifically involving YouTube‘s recommended videos, which are indirectly related to video content. This thesis investigates the problem of predicting a video’s popularity (measured by the number of views over time) using directed graphs constructed from sets of trending YouTube videos (split by genre) and random YouTube videos, due to the vast differences in viewing patterns for videos of different genres. Furthermore, datasets with different constraints on the videos included allowed us to compare the effects of videos with varying degrees of popularity. The pattern of voting on the videos (using data on likes and dislikes over several time intervals for each video) is used as an a signal for the quality of a video, giving further insight into the relationship between a video’s popularity and the quality of the content as assessed by those who view it. We see a small-world phenomenon occur with most of the recommendation graphs, indicating that recommendations tend to remain within similar groups of videos. For predicting popularity, we find that the Hawkes Intensity Process works particularly well for this type of data, and that clustering and community detection techniques that group videos based on similarity do not necessarily lend themselves to predicting popularity.