Information flow reveals prediction limits in online social activity
blog.acolyer.org
blog.acolyer.org
For example, they excluded retweets but what about replies? When your friend asks you about [TOPIC] in a tweet, of course your reply is going to be related to [TOPIC].
What about Twitter users who mostly use it to comment on breaking news? Of course 90% of the words are going to be predictable, but it's the 10% that are different and important.
[1] - None of their charts have y-axis starting at 0, leading to a much more extreme-looking difference in results. https://arxiv.org/pdf/1708.04575.pdf