Not a direct answer to your question, but at least in English there isn’t that much stuff on how TikTok’s recommendation system works internally. This is the best breakdown I have found on TikTok’s recommendation system internals: https://leehanchung.github.io/2020-02-18-Tik-Tok-Algorithm/
TikTok seem to be learning from what the user is actually watching and for how long and not just the user's "Like"/"Not Interested In" actions. However it still seem to learn from the "Not Interested In" action more than any other platform.
This is a pretty misinformed take when it’s publicly known that YouTube was already doing this (learn from what the user is watching and for how long) the year Bytedance was founded (2012):
Somehow they're doing it better. At least subjectively, people complain more about the YouTube algo's performance than tiktok. For the latter, the most common complaint is that it's too good.
Does YouTube actually work for someone? I bit the bullet after not really giving any other input than what I watched and started marking "Not interested" or "Don't recommend channel" but I cannot see that it has had any effect at all it keeps recommending rubbish "sensationalist" videos.
I know that they probably optimize for ads etc. but if they actually showed me videos I would like to see, then I would spend more time on the platform.
These sensationalist videos are a plague for me. It only takes watching two or three regular videos on a specific topic for YouTube to hone in on it and start populating my feed with them. I find it very disappointing that they exist in every area that I am interested in, including relatively niche ones like functional programming and mechanical keyboards. Although, I have yet to come across a sensationalist video about prolog ;)
I have the exact same experience. I want to spend time on the platform and I can find videos that I like and would have loved for the algorithm to find for me, but I really have to search very thoroughly for them myself.
Yes, it works great for me. I don't really use the negative feedback. Most recommendation I guess are new uploads from creators I like or videos I haven't watched from creators I like. It shows me videos from unknown creators to me that match my interests as they go viral. It shows some random old videos that just randomly go viral again. It's a good experience.
That's interesting and perhaps I should not be that surprised. If it did not work at all, then they would probably have changed it. Their KPI is probably ad revenue, but still if it worked for no one, then I would expect their revenue to decline as people spend their time elsewhere. But then again, where will you go and watch medium to long videos created by "normal" people? Vimeo? Perhaps youtube's interactions are despite their algorithm and not because.
Not directly, most people believe they optimize for session time. It tries to serve you a result that will keep you on the platform watching videos as opposed to needing to keep scrolling or leaving the platform. They truely do want to serve the best videos that they can to you that they think you will be interested in watching. Thankfully for YouTube you watching videos and YouTube getting ad rev is correlated.
>where will you go and watch medium to long videos created by "normal" people?
No one forces you to watch that format of video. You can go on TikTok, Twitch, Netflix, etc and be entertained.
Technically, yes, but when they're talking about this sort of thing, they mean "personal recommendation system" or "content-based recommendation system."
For example, the HN front page is a recommendation system if you literally mean system-that-recommends-web-pages-to-look-at. But it's not personalized; every visitor sees the same front page. This fundamentally makes it a different sort of thing.
If You have 10000 posts that You have to sort it in some way and the user just going to see 20 of those, the sorting is the recommendation system, people are just used to think of products, movies and songs, but in those platforms the users are the products
This hardly seems like a reasonable way to characterize Netflix, which has a personal recommendation system, especially compared to HN, which is ad supported yet gives the same recommendations to everyone.
My reading of their comments is that they are trying to say that social media and news media can be characterised as having recommendation systems too, not just song and movie platforms (I don't know who exactly they're arguing against – I've never heard anyone say that recommendation systems can only be for songs and movies).
I don't think they're really paying much attention to the dimension you're splitting it along, i.e. whether the recommendations are personalised for each user. The huge important idea they have in their head is that recommendations can apply to user-generated social media content too.
Yes, but this is an extremely important dimension to split on. The practical implications are so big that the game changes completely.
* HN and classic Reddit sort their items on a single dimension ("hotness"), calculated using a few input variables and producing a single output variable. This is about as cheap to calculate as recommendation systems get. The XKCD comment recommender is a bit more complex, but still in the same complexity class. Since the whole point of an algorithm like this is to be timely, the naive approach is to compute it on-the-fly, which it's perfectly simple enough to manage.
* At a somewhat more complex level, you get stuff like a basic, uncustomized Similar Items list. If YouTube has no data on you, this is what you get from their sidebar recommender (and their front page would be analogous to Reddit and HN, but sharded by region and language). It's also pretty close to what AdWords used to be, before they started doing user profiling. The thing with this method is, even though it involves some level of AI, it's presenting the same thing to everyone and it's expensive, so the natural solution is to precompute it.
* Personalized recommenders are the worst of both worlds. You can't naively compute it on-the-fly, because it's too slow, but you also can't naively precompute it, because there's a combinatorial explosion of users and items. You actually have to be clever about it.
