[ryanbrunner]

I think the fact that most recommendation algorithms have seemingly converged on what seems like a really poor and naive implementation - fixation on very recent activity - shows that the sort of deep personalization touted is mostly BS.

Both YouTube and Amazon heavily personalize by recommending primarily the 3-4 things that I've interacted with in the very recent past.

[charcircuit]

This is not true. For example every time Summoning Salt uploads a video, which happens every few months, it will show up on my recommend feed because YouTube knows I'm willing to watch their ~1 hour documentaries even though I'm not subscribed to them.

[Jweb_Guru]

Youtube seems to be a rare exception here in that people actually feel like its algorithm is useful. However, even then, their algorithm mostly seems to devolve to "what creators have you usually watched videos from" and (usually directly after you watch such a video) "what videos did other people who watched that video watch?" Basically the same principle as PageRank, just with a lot less spam to deal with.

[sdoering]

This could (probably isn't) be a very quick implementation with a heuristic like 'if percentage of viewed videos from channel x (essentially per channel viewed) > threshold ==> show new video from channel x on homepage next time user appears.

Make it fancy and use a multi armed bandit and call it machine learning/AI/data science.

[earth_walker]

What it proves is that despite all that personalized data they have, it's the naive implementation that gets them the most clicks per dollar.

So the question is this: if they're not (and never were) using that data for what they say they were, what are they doing with it?

[fomine3]

I believe YouTube recommendation is most well working one, so some people getting into echo chamber.