how proper are we talking? "for you"-like?

in that direction - certainly a level above regex-y ones. Like I don't really comprehend how you're supposed to figure out what's relevant, like I could imagine how you'd do it for one specific person but figuring it out on the fly for everyone who might possibly look at your feed? no clue

me, i would: a) restrict myself to one language at a time b) keep one global model for all users in that language, and keep one going for each subscriber c) build a huge table of {user, post, bools for liked and reposted} d) define Happy Points scale: 1 for a like, 10 for repost

e) generate the global model aggregating over all users, and predict from post features to Happy Points. I would start with a random forest due to easiness: scikit-learn.org/stable/modul... f) do the same for the user-local models, except only look at that user's rows in the Big Table

g) predict Happy Points as a weighted average of Global and Local predictions, with weight accruing to Local as you have more rows in the Big Table from that user. h) Each refresh, rank all recent never-before-presented posts by predicted Happy Points. Log what youve shown to build up the big table!

i) finally, check which posts got likes/reposts, get that feedback into the Big Table from there -- that's it. tweak the knobs around (how recent is recent enough to rank? how should Global-Local weight shift with user row count?) in randomized A/B tests of different settings

ive totally omitted how you turn "details about the user" and "details about the post" into the array of float that the regression framework is going to want. let me plug my company and say here's how you can convert natural language to such an array: cloud.google.com/vertex-ai/do...