Hacker News Reader

Comment by HardwareLust | Hacker News Reader

YouTube doesn't seem to understand my preferences, at all. It's far, far too biased toward what you recently watched. It basically doesn't work and certainly doesn't "learn" anything.

View on HN

YouTube absolutely IGNORES your preferences.

I watch it daily, I have never listened to any hip hop, I don't watch sports nor trending topics, yet every single day it suggests football videos, MMA, the latest hiphop hit, and celebrity crap (imagine the Kardashians) among my other interests.

I click "don't recommend this channel", or "I don't like this video " every single day, I have an extension to make this convenient so I train it dozens of times a day yet it doesn't bloody care.

I pay for Premium because I watch it on my iPad and smart TV, but I'm planning on telling Apple to fuck off, buy Android everything and sideload a system wide adblocker so I can stop paying for that terrible service.

This is the inconvenience it forces upon a paying customer. What an abusive relationship.

https://github.com/yattee/yattee

This may be in your interests if you have an Apple TV.

I think it supports iPad, as well.

Thanks. I should really dedicate a weekend to setting something like that up on all my devices. I'm willing to put up with bad recommendations if I know they don't get a cent out of my views.

Sort yourself out, Google.

For me it recommends exactly what I watch - I've never seen any recommendation for stuff I don't watch.

The trouble seems to be it can't seem to find anything new. So if anything, I'd want it to branch out a bit and find semi-related stuff, but it doesn't.

You're not the first person to complain about YouTube recommendations, and I don't know why my experience is so different than yours.

My experience is the same as yours. I wonder if there’s some giant A/B test going on.

Somewhere in the middle for me, but what it recommends to me are generally channels I am already interested in. I think the only novel recommendation I've seen recently, that I enjoyed, was of some kind of classical Chinese music. Otherwise, I see a lot of absolute fucking garbage from the likes of Mr. Beast.

It's insanely biased towards popularity. I watched this series of like 40 videos on extremely large numbers. And not one time did I ever see one in my recommendations. You would think when you've watched 1-20 of something the algorithm would be smart enough to suggest 21 but no.

It does do that for me, and it also regularly recommends videos with <1K views if it's from a channel I often watch. It even sometimes recommends videos with <100 views from channels I don't know if it's on a topic I watch videos on regularly. So it seems the algorithm doesn't do the same thing for everyone.

What's funny is that the super smart AI behind the rec engine in YouTube sometimes is so dumb it can't figure out that video #3 in a series with consistent titles should be followed by video #4.

Sometimes it just gets it wrong and skips one, other times it just recommends some other totally unrelated video even if you've been watching the same numbered series forever.

And it's got a short enough memory span that one day it'll recommend out of the blue, for weeks, video #57 of a series you've seen a couple years ago.

I wonder how many of the "algorithm's" problems are caused by a huge number of literal toddlers that just hit buttons at random. How many parents just sit their kid at an iPad for some time to do something?

Make sure your watch history is enabled, without it the recommendation system cannot learn.

My watch history is enabled, because I've learned that if you disable it or ever clear your history, you're back to square zero.

> I watch it daily, I have never listened to any hip hop, I don't watch sports nor trending topics, yet every single day it suggests football videos, MMA, the latest hiphop hit, and celebrity crap (imagine the Kardashians) among my other interests.

I don't have this experience at all. YouTube recommends me pretty much exactly what I've been recently watching. I watch a lot of long form interviews and tech videos, and my recommended feed is entirely filled with more of the same, which are of decent enough quality for me.

I wonder if something is wrong with your account?

I used to have the same experience of YouTube recommending just more of whatever I just watched, and at most recommendations would stick around for 6 weeks before it forgot about a subject. That was until earlier this year when I watched some videos from MrBallen (camp fire stories based on real events), and ever since then I have gotten my recommendations filled with crazy conspiracy stuff.

It stopped recommending his videos months ago, but lasting damage has been done to my profile or something like that because the conspiracy stuff sticks around and has crowded everything else out. The old behaviour of fixating on something I just watched for 6 weeks no longer happens.

I suspect that certain types of videos/subjects are able to leave a massively oversized weighting for a category on your profile that everything else get's completely dwarfed by.

Out of curiosity, do you watch those videos? Is that what trains the algorithm to show more of them?

Not parent but same experience. I watched MrBallen's videos as "creepypasta based on real events", and I guess that is a pretty similar vibe to conspiracy theories. Never watched any of the conspiracy content and still got it in my recommended videos for a long time.

Nope, never watched any of the conspiracy videos. I did however watch quite a few of MrBallens videos, specifically all the cave/diving ones, they are just great stories.

Exactly the same experience here. I kinda like MrBallen's presentation and I know I can't take his stories at face value, but apparently it's so close to conspiracy stuff that it just overwhelmed my recommendations. I guess part of the problem is that conspiracy content has very high engagement among its audience, so youtube is biased towards it.

Ultimately I had to go back and scrub his videos from my history and even then it still took a lot of effort saying "not interested" to conspiracy stuff until I no longer got those suggestions.

If you want to fix it (and I agree it shouldn't be this painful, but you deal with what you got), you can go back in your history and remove the videos from MrBallen from your watch history.

This should fix it for good, almost immediately.

Thanks

> YouTube recommends me pretty much exactly what I've been recently watching.

It depends. Maybe I wasn't clear, most of the recommendations are fine (not great, but still pertinent), but sections like "New to you", or "Just uploaded" or even the search suggestions are full of trending videos that I keep telling it I don't care about.

Right now I just get recommended three categories of videos:

- 90% relevant to my interests but often stuff I've already watched. Nothing that strays a little further than the channels I already know. Keeps recommending stuff I've watched and liked months or years ago. Video #42 of a series I've seen fully.

- 9.9% useless trending bullshit. This month it's all about the "Why have we divorced" or "Meet our new baby" videos. I don't have children, I'm not even in a relationship. Google decided I would care about this.

- 0.1% some genuinely new stuff that's relevant.

If I remember what I did correctly, you can block most of the named sections using something like uBlock.

> sections like "New to you", or "Just uploaded"

I actually expect those sections to completely ignore your preferences. They're just advertising sections, so they reflect what YouTube wants you to watch.

I got a lot of recommendations for stuff I ALREADY WATCHED, it even shows the red bar showing it KNOWS I watched it, still recommends

Youtube does that because people like me compulsively re-watch content. Youtube probably needs a few user-visible knobs like "I like to rewatch stuff vs I hate rewatching stuff"

> I wonder if something is wrong with your account?

If so, the same thing is wrong with my account, and that of my other friends who watch YouTube often.

Do you perhaps have some account setting where it doesn’t store info or track you? In other words, are you preventing them from understanding you?

I have a completely different experience. My recommendations are pretty decent, if recent-biased. You’re not on the “explore” or “trending now” page right? Obviously trending videos aren’t tailored to individuals

Sounds like you have a child or relative who might be using your account

I feel like every recommendation service works this way - YouTube, Amazon, Spotify, etc. Whatever the last thing you looked at becomes your entire life's mission according to these systems. When I think about all the buzz words around machine learning, ai, and so on, it's completely incompatible with the reality I see. What exactly are all these engineering teams doing just to end up with this result?

Amazon might be the worst example though. Buy something like a bike rack for your vehicle and Amazon will recommend nothing but other bike racks... as though I need dozens of them for one car.

> Buy something like a bike rack for your vehicle and Amazon will recommend nothing but other bike racks... as though I need dozens

Perhaps they beta tested only with Jeff Bezos who had this exact quandary.

Youtube does have a great escape latch though: go to the History tab and delete your viewing history. Your recommendations start from scratch and you’ll get funneled into a potentially new corner of videos.

With that in mind, I think Youtube recommendations are quite good. Just do it when your recommendations get stale, monthly or so.

I wish other platforms had that like Amazon.

Amazon actually has a pretty detailed page that lets you control what feeds back into your recommendations.

https://www.amazon.com/gp/yourstore/iyr/

Thanks, I had no idea. This is great.

It'd be great if I didn't use my History to actually find videos I've previously watched...

I don't understand why recommendation engines just don't provide a reset button. It'd be better for everyone involved: content provider, content creators, users, ad buyers.

> With that in mind, I think Youtube recommendations are quite good.

Having to nuke a valuable reference like your history in order make the recommendations approach something useful is the opposite of "quite good".

Well, I'm one to appreciate what we have available to us especially compared to the alternatives and temper my appreciation accordingly.

I don't know if I use a single service that gives me particularly compelling control over recommendations. Though I also don't find myself deep-diving my history so often that clearing it every few months affects my life, but I suppose I can see how that could be a setback.

My main point is that if the recommendations are improved by giving YouTube _less_ information to work with, that's a solid sign that the recommendation system is borked.

Amazon really, really likes to recommend me a wide variety of shower curtains after buying one, 4 years ago.

Give it a rest already.

They realllllly want me to buy far too many vacuum cleaners for the one rug in my house :|

With Amazon quality control, I’m sure you’ll need a new vacuum soon enough.

Unfortunately Tiktok is the only platform that nails it for me. It regularly manages to find old hobbies or interests back into the rotation,

Youtube’s recommendations are just matching my most recent searches while occasionally throwing in Nazi crap I never asked for.

It appears to any end user that knows some basic logic and looping concepts that the advanced AI/ML/whatever recommendation engine is no better than looping over the list of recent content (videos, products, etc.) and selecting more content that has same keywords. If you remove the volume and distributed systems elements for large systems its probably a single SQL query joining 2 tables, history and products (or videos or whatever).

For the bike rack example, to make it "smarter" you could have a product type that is something like "one-time purchase" and have a linked product type that is accessories for the one-time product, like helmets, pumps, tire changing kits, inner tubes, water bottles, etc. This would be what a retail worker at a bike shop would do - ask if you need any accessories, maybe point out clearance/sale items. Thats it, no fancy engines/AI/ML needed. If the customer wants other stuff, there is a search bar they will likely use.

I chalk it up to Resume Driven Development and just the innate tendency of engineers over-engineering things because that's how our brains work.

Most preferences type systems don't seem to get the nuances involved.

I just watched a pleasant video about a particular video game.

Youtube now wants me to watch the most obnoxious youtuber they can find who also covers that video game.

No, actually the reason I watched that other video was because the person in it was nice and pleasant.... the recommendation is the LAST thing I wanted.

YouTube doesn't care about your "preferences." It only cares about what it predicts your behavior will be. If you tell your drug dealer that you're trying to quit crack, he's still going to sell you more crack.

Well at least it doesn't work like Instagrams where it doesn't do anything at all. I don't want to see hundreds of videos making videos of street vendors all doing the exact same thing.

Yeah, on a non-logged-in system I watch YT on, I watched exactly ONE gaming-related video, and now suddenly the "home" view is full of Minecraft, Roblox, whatever FPS shooters, etc... Like... I'm not suddenly absolutely obsessed with a topic because I viewed one video, guys...

It operates similar to saying to a spouse or in-laws or other family that you like some <collectable item> and they are the type that runs with it until you have to awkwardly tell them "No more". Every holiday, birthday, random "oh I saw this and thought of you" type of thing.

this is my experience too, as well not showing videos from people you've subscribed too but maybe haven't watched in a while.

the most shocking thing is they still have not managed to get a decent recommendation engine after all this time.

Reply on news.ycombinator.com