Ask HN: Why YouTube suggestions are better when I am logged off? | Hacker News Reader

Ask HN: Why YouTube suggestions are better when I am logged off?

24 pointsonurcel23 comments

I watched a few videos on youtube without being logged in, and to my surprise, suggested videos were much more interesting. When logged in, I mainly get more or less related videos I already watched or videos with a lot of views.

What are your thoughts on that?

View on HN

They really should stop showing videos from your subscribed channels in the 'Home' view. They take ay too much priority there.

If you're finding that YT, when logged in, is becoming obsessed with any one topic, you can go into your YT history and remove some of the entries related to that topic.

That'll remove those from consideration and your home screen recommendations will change.

It has gotten to the point that I have trouble discovering new interesting things -- unless I already know what I'm looking for I probably won't find anything to watch that I haven't already seen.

I clear my search and watch history every few weeks, sometimes more often. It does improve things when you notice that everything in your recommendations is stuff you've already watched or very closely related.

Id wager its placebo. I have YT history cleared and disabled for years now, same recommendations as before that.

I've found the same. I think they focus too much on what I'm subscribed to even on the "home" page vs what i've been watching/searching for in my experience. I now use incognito a lot, except the nag screens for signing in or youtube tv have gotten ridiculous...

Agreed. I find the Youtube suggestion more useful to me when I log off. I may start searching videos of interest and get suggestions of related videos. After it's done, I just clear the cookies. It's like building a new profile of minimal necessity every time.

>videos I already watched

Wouldnt be the end of the world if they didnt screw with CSS to hide that fact. One of the best simple mods you can do for the sake of your own sanity is injecting those two styles into YT website:

    a #video-title { color: red !important;}
    a:visited #video-title { color: black !important;}

Right now opening new Andrew Camarata video from my Subscription subpage shows me 20 recommendations total:

16 old Andrew Camarata videos I already watched in full

one playlist of all Andrew Camarata videos ...

two new to me videos of other people doing related things (TILE PLOWING, fixing a trailer)

one new video related to my overall subscriptions

This is pretty much the norm on YT right now, 3 new recommendations, 17 videos you already saw.

Cosine similarity. Recommendation is a search problem. YouTube has strict latency targets. The results are skewed toward what is easy to return. Videos already watched, videos from your subscriptions, and videos on topics that dominate these are easy to calculate and cache links. Deep insight is likely to time out most of the time.

Why can’t that insight be done offline?

I would bet it is to the degree it makes business sense. But there’s a lot of other business considerations.

This isnt why and I'm not sure why you typed it up with such confidence

I am not a fan of Youtube's recommendation system either. I think it favors too much "more of the stuff you recently viewed". Sometimes I watch 4-5 videos on a subject, then I move on but my feed keeps contaminated suggesting more about those things over and over again.

Youtube should really trying recommending stuff that's not "exactly the same as you saw, but from a different channel".

Also, for some weird reason (and partially contradicting what I said above) Youtube has been insisting in suggesting Jiu Jitsu videos to me for a very long time. Sometimes I watch MMA-related or Muay Thai-related stuff, but never Jiu Jitsu stuff. And it insists on always showing me Jiu Jitsu stuff.

Anecdotal evidence warning:

Basically it's trained to feed you what it thinks you want to see. From my experience, it doesn't matter how many different channels you are subscribed to, it looks at metrics from the last hour or so of watch time and recommends based on that information. You end up silo'd in a small set of refined options based on every video you watch.

The best way to get a fresh perspective on Youtube is to logout or go to the trending page ... otherwise you're sitting in an echo chamber of only things you agree with... which isn't that interesting

I remember a few years ago the Amazon recommendations where like this, not sure if they still are. "You just bought a blender, here are 5 other blenders you might like."

There are two types of familys for recsys

User-item and item-item, recommendations are often a mix of both, yt in particular has a skew to use user-item based on your history ( maybe it's because they found out that people like to view the same thing or more from creator they already know) but when the mix is skewed liked that you tend to miss on de serendipity of the recsys

It all depends of the target metric that their models where trained, if you look at their papers they said that they want to maximize view time

Because Youtube's algo is trained on videos you are most likely to watch through.

None of which really take in account your personal preferences. Youtube only cares about how long you stay on a video after clicking on it.

https://www.wired.com/story/youtube-algorithm-silence-conspi...

Slightly off-topic, but it's really annoying that the YouTube Android app made it basically impossible to log out. Even the workaround to reset the app to the factory version and never logging in doesn't work any more. Now the YT app will automatically log you in with your system-wide Android account. So you have to log out from the system account and lose features like syncing.

You can just use YouTube Vanced or NewPipe for that

I would speculate:

logged in you get the best videos from a pool of narrow interest, which you have probably already mined for the top videos for months or years.

logged out, you get the best from a much larger pool, which probably gets new material faster and you don't watch often.

Not my experience. Collaborative filtering makes it that it recommends what other users are watching (primarily based on location). Let's say that it reminds me that I have almost absolutely nothing in common with my "compatriots".

Same! Why is this? I watched some videos on car stuff and the next day I get suggestions for people unboxing meal planning stuff when I never watched a video pertaining to that.

because algorithms can't REALLY predict our preferences?

I think it's that they work OK at first, but then fall into patterns. Like getting stuck in a local maxima.

It also might be true that we can't predict our own preferences! That could explain why we justify our choices after the fact - deluding our conscious mind to thinking this was what we wanted the whole time.

Reply on news.ycombinator.com