The You Tube Algorithm is largely a black hole. One that seems to work mostly based on what it thinks you want. Or, at least that is how they would like you to think that it works. But, even if we take it in the best way it is trying to figure out what you are most likely to click on initially, and keep you watching for as long as possible.
And how does it determine that?
Given the available data it seems to rely, at least in part by what other people who watch the channels that you watch will also watch. It makes a profile of the typical user, and likely customizes that based on the interactions between the typical user profiles.
They Don’t Know Why A Person is Watching
I might be watching a baking video for a special event, or put on a music channel to sooth my dogs during a thunderstorm instead of using Spotify. Recommending me things like those items is likely to have very little relevance to my wants and likes. (Seriously, I did the dog thing once in 2018 and now I can’t get Low Fi (and similar BS like Vevo, YouTube Music and artist channels) to go away. 1, bloody, time.
And for better channels, it may not even be the type of content at all. For example, I don’t watch Markiplier because I like lets plays. Yes, I’m a gamer, but mostly I just like Mark. I found his videos at a bad point in my life, and he is a kind, engaging, funny person. It’s his personality that I’m drawn to, and admittedly some of his very cowardly overreacting screams can make me smile on a bad day.
You can suggest the rest of gaming You Tube to me all you like, but I probably won’t watch it. I don’t care about them. If you can’t identify why audience segments are watching, and which segment a viewer belongs to, then you can’t really make a decent recommendation based off of it.
Basically, its using code to create the assumption of a false equivalence, a real problem in AI development and shows that the developers need to think a lot more about their propositional logic before they can call this part of the algorithm even close to a beta version.
It Homogenizes The User Base
I get lots of sports content recommendations. Just one problem. I don’t care for sports. The same with news content and all other content from traditional media content channels. I don’t want it from YouTube. But no matter how much I tell it I do not want those things it ignores the feedback. Because most users will accept this content (and similar content from so called “real” celebrities) and let it run, we are all stuck with it.
In trying to be all things to all people, the system ends up seeing no people. It makes generic the user base to a point where it assumes the lowest common denominator. Segmentation is important for solid predictive algorithm development, but this generic approach is stubborn.
I’ve listed ‘Not Interested’ and “Don’t recommend this channel’ for dozens of sports-related media by now. If it had any kind of decent segmentation system it should have filtered that out by now. The same goes for broad gender/age categories (outside of under 18) because by now we should all know that those are far too broad to be useful.
It Creates A Feedback Loop of Viewership That Limits Channel Availability for Users
This is legitimately two full rows of the recommendations on my home page. Its all the same channel. So, if I take risks and try different things (as established above) the algorithm will punish me by not letting it go. If I don’t I get this crap. Its a lose-lose proposition.
It Creates Recycled Content Farms and Strangles What Most of Us Came For
I don’t care about network TV content, or things that are the kind of evergreen bullshit over and over again. There has been a rise of channels that just repeat the same things over and over again (a la 5 Minute Crafts or Mr. Cake) that make and re-stitch the same junk over and over again and get rewarded.
Your users didn’t come to you for the same crap recycled, or to see Jimmy Fallon (seriously, the man is annoying AF and the pathetic attempts at #relateablility that have been brought on since quarantine started are asinine. Keep your brats off camera!) and his late night cavalcade of 1990s jokes.
I want to support real creators, not polished bullshit brands that are making the platform mostly unusable.
It Ignores User Feedback
I can’t count how many times I have told the system I do not like a video or a channel, and three refreshes later (I did a bit of an informal experiment and this was the average number of times) it brings back the same thing I just told it I don’t want. Oh, and it still gives those results high placement even if you told it you don’t want the channel.
For example, I told it ‘Don’t Recommend This Channel’ for Netflix, then a while later I searched a broad term (Halloween, if anyone is curious) that should have had millions of results, and what’s in the number 2 spot, Netflix. Of course it is, because its ignoring the feedback it was given. Likely because the channel is popular.
But the bottom line is that I don’t care what’s popular. I care what I like. And deciding to lower weight user feedback in favor of overall popularity tells the user that you don’t give a drop of care what they think.