Hacker News Reader

To add on to your point:

Given my understanding of how the recommendation engine works, if you watch a video slightly related to a topic, then another video slightly related to that topic, there is a reinforcing effect whereby the engine believes you to be more interested in that topic. E.g. if I watch a video on vegetarianism, then another one, it may strengthen the confidence that I am interested in vegetarianism, and conclude that they should be recommending me more content about vegetarianism and perhaps even content about veganism. Apply this to something like conservatism and you'll be getting recommended alt-right content in no time.

Slightly off topic:

I have a theory about this which is that there are certain recommendation loops or "attractors" that you basically cannot escape without manually flagging videos as "not interested", disliking, manually choosing an entirely different topic, or taking many days off of the platform.

I also am curious to know what percentage of content literally never gets recommended. I've been down the rabbit hole, so to speak, on certain youtube topics and I get the sense that at a certain point, there are many videos on the topic that youtube simply isn't recommending. E.g. if I'm watching a video about the Philosophy of Derrida, I feel like my recommendations never show lower view count videos related to the topic, they are much much more likely to recommend higher visibility channels full of slightly related content.

> if I watch a video on vegetarianism, then another one, it may strengthen the confidence that I am interested in vegetarianism, and conclude that they should be recommending me more content about vegetarianism

Recommendation algorithms don't "conclude" anything though, they are merely statistical tools to ensure you get somewhat relevant content, but they are never perfect because there is a lot of noise in what everyone watches.

Not perfect is understatement They are horrible. Not just because of what the article is about, but because their recommendations are ridiculous avwn outside political content.

And youtube rules also motivate providers to churn out a lot of content regularly (leading to quick to produce crap) and punish those who take their time to think or research before they talk.

>Not perfect is understatement They are horrible.

They would be much less horrible if every time one watched videos, they would rate them in a way, instead of assuming that "watching"="interest".

> And youtube rules also motivate providers to churn out a lot of content regularly (leading to quick to produce crap) and punish those who take their time to think or research before they talk.

That's not just Youtube: about all media is driven by what's "new" and "hyped" rather than what is deep and well thought about.

Watching does = interest if you continue to stay on the topic. Interest certainly does not mean approval, and it often can mean "I find this content extreme and outrageous."

I am interested in topic of WWII including seeing Nazi movies. I am not interested in nazi stormtrooper adjacent alternative history channels youtube recommends to me as a result.

I am interested in scientist or writer etc youtube channel about her topic. Not so much in confident-Johnny-cranked-out crap on the same topic.

It becomes much less horrible when you consistently train the algorithm with the "I'm not interested" button. You can even choose if you're not interested because you already saw the video, dislike the theme, or dislike the channel. My YouTube home page is pretty good now, after doing this about 20 times. Any time I watch something vaguely conservative and start getting "SJW CUCK OWNED!!!" everywhere, one button press is usually enough to instantly revert back.

The problem is probably 1% or less of users ever click it.

Recommendation algorithms don't "conclude" anything though, they are merely statistical tools to ensure you get somewhat relevant content, but they are never perfect because there is a lot of noise in what everyone watches.

A. Jeesh, there's no problem with informal anthropomorophizing in this situations. When humans have a goal and feedback towards reaching a goal. When a human gets positive feedback that X gets them towards the goal, human choose X. The combined system Google-corp+developer+algorithm is also goal seeking and making choices so anthropomorphizing the system is appropriate.

B. The problem we're talking about isn't "noise" but "feedback" - a goal-seeking-system that muddies it's final result with it's initial state. Essentially, bias, a situation that's quite common in statical systems.

I have a theory about this which is that there are certain recommendation loops or "attractors" that you basically cannot escape without manually flagging videos as "not interested", disliking, manually choosing an entirely different topic, or taking many days off of the platform.

I'm pretty sure this is true. However, there is one more effective around this. You have to go back and delete your history. I have one youtube identity I maintain for exactly one interest (call "interest X"). I have to essentially scrub everything not related to that from it's history or more and more unrelated crap will appear reliably. Basically, even watch 10% of something means recommendations jump up to 50% of other things even though I have a looooong history of just being interested in interest X.

I see the same thing on a completely uncontroversial subject: music videos. If I watch e.g. a Beatles video, the recommendations aren't [Rolling Stones, Kinks, Badfinger, ...], they're always [Beatles, Beatles, Beatles, ...]. I don't know if it's naive and bad or if it's actually good for engagement and I'm just an outlier in what I expect to get.

I only wish I would get songs by the same band... It sort of seems like Youtube tries its darndest to steer me away from the current band. I'd be happy if it just cycled between hits from the same artists, but I've never seen it do this recently.

I think what's most frustrating about it is I don't know what game they're playing. Recommendations used to be logical, now they're almost a waste of screen real estate. What even happened?

Dislikes and "not interested" are important.

I've noticed the same in my usage across almost all google services at this point. My speculation is that this is due to them now being so heavily dependent on neural networks these days vs the previous mix of algorithms. The old approaches seemed to be better at not presenting false positives / creating a rediculous mish-mash of previous trains of thought in an effort to recommend something, anything... no matter how wrong.

For example: older versions of the google keyboard were pretty good about not recommending/autocorrecting non-words. So I'd type things like function/variable names (which often use camel case) and it would either suggest valid English words or perhaps a non-word used earlier in the message. These days, my use of camel case names in technical emails spills over so that emails to family now pop up with bizarre camel case recommendations. I seeing similar things happening to YouTube recommendations and elsewhere.

My guess is the system creator think a lot about making the best recommendations for situations where they have enough information but seldom think much about the consequences of making recommendations where they don't have the information to make a recommendation and both users and developers underestimate how information is needed for a decent recommendation.

Edit: And the situation seems very much a product of youtube (and all the content providers, basically) stripping any user content controls from their UI. The user has no easy way to tell youtube what they want besides choosing between video that youtube offers them and those become more circumscribed as the process progresses.

There's a myth of the content providers doing magic to give the user what they want and the results seem to be this, ie, just terrible.

For episodic content it actually works ok, recommending videos on the same series, even if it's not the same channel

Ideally it would know how to predict order from title but that doesn't always work

There's is a hidden "not interested" button on mobile where you can then say why you don't want to watch a video and choose to block the channel or similar videos.

The algorithm isn't ideal, but after a couple days of filtering you can pretty consistently get a feed of videos you're interested in.

The idea of people getting persuaded by fake news, etc. is a topic for another day. This assumes you can vet trustworthy sources on your own.

I've noticed this as well, the algorithm hones in on a few topics and videos. Interestingly enough, so does the algorithm on the Quora app. I hate this behavior, I want to use a recommendation algorithm to discover new similar content. My guess is that a "new similar content" algorithm doesn't get the same levels of engagement. Maybe its the same as radio stations playing the same twenty songs on a loop.

Clearly google knows what their doing, its just disheartening that this algorithm is the best for engagement. Also makes me wonder about human habit and reinforcement...

Here’s a Quora response to the question of why radio stations play the same songs over and over and over and over and over and over and over and over and ... again:

https://www.quora.com/Why-do-radio-stations-play-the-same-ha...

I am fond of videos showing people's trips in Antarctica. This tendency exposes me to flat-earth videos ALL THE TIME. No matter how many pleasant tourist videos I watch, I still get flat-earth videos and even occasional hollow-earth videos. (the flat earth view includes the notion that the Earth is a disk with Antarctica as the outer edge).

I can't watch the pleasant tourist videos of Antarctica before bed, in case I fall asleep and the automation shows me a stream of flat-earth videos.

The link here is just the word Antarctica. The flat-earth view has nothing to do with viewing penguins or pleasant scenery.

Another aspect is youtube's algos tend to heavily favor "engagement" ie watching a full 30 minute video, not hard to see how this funnels people into fanatics due to their nutty followers devoted engagement.