I make algorithms that filter RSS feeds and other social media content. You don't need a global view of the system at all to do this, you just need
a point of view. That is, if you have a few thousand posts and a thumbs up/thumbs down judgement you can train an ML model that will predict those judgements.
With about two days looking at toots I could make a model that shows you nothing but angry toots about politics or one that removes angry toots (could take down that keyword filter that means I never hear if somebody is having trouble with the transmission in their truck or that Transnistria got invaded.)
The main reason I haven't developed a social media (as opposed to product/service) sentiment model like this is that it would involve looking at a few thousand angry toots. (1) The reason I want it is that I don't want to read those toots, (2) it would cause me great suffering to look at those toots. Social media moderators at companies like Facebook have been traumatized, it's no joke.
If it was my social network, I'd use that filter to put a brake on selfish angry memes spreading so that the pain of one person reading angry toots gives relief to so many more.
I have a model that predicts the probability of a headline getting a lot of comments relative to votes on hacker news: some high scoring headlines in my RSS reader right now are:
Why do women commit far less crime than men?
Study suggests anti-Black racism may account for conservatives' negative reactions to jobs requiring DEI statements
Checking a bag will cost you more on United Airlines, which is copying a similar move by American
Everyone seems to forget why GNOME and GNOME 3 and Unity happened (2022)
Three of those are clickbait, the last one is a good HN submission. A social media sentiment model can give a larger algorithm a "superego". There are other ways to pursue engagement other than selfish angry memes.