Live data from Hacker News

On YouTube’s recommendation system

blog.youtube

81–90 of 337 posts

Re: On YouTube’s recommendation system

#81
"videos from reputable sources like Vox" - yeah, right...

I wonder who that article is addressed to? At first it may seem they want to reassure viewers that they will get shown the best content. But the longer the article goes on, the more it sounds like a reassurance that they will absolutely block out unwanted political views.

Why do they feel the need to put out such an article? To stave of government regulation? Or for their own political activism? To reassure the leftist mob that they won't be part guilty of electing another Trump?

I mean as a user, I'd like reassurance that the algorithm tries to show me videos I want to see, not videos that others want me to see.

I am not afraid of being shown a flat-earther video, so I don't get the appeal of the promise to not show me certain things.

Re: On YouTube’s recommendation system

#82

For me, it seems like youtube's recommendation system mostly gets stuck in the local optimum. Once you watch a video from a particular channel, it'll keep on recommending the videos from the same channel for months. Soon, the recommendation just consists of stuff from the channels that have you already viewed. You rarely see new types of recommendations.

It seems to me that YouTube is doing complex bucket testing with the recommendation algorithm. Some days, my recommendation list is full of obscure-but-relevant-to-my interests content that would require legitimate ML to populate, and then the next day my feed is basically

  SELECT DISTINCT videos.* 
  FROM videos, followers 
  WHERE videos.creator=followers.creator
  AND followers.user = $ME

Re: On YouTube’s recommendation system

#83

Why don't they let me tell them the interests that they are trying to infer? I'd be really happy to review sets of recommendations and indicate the kind of videos i like and the kind i don't.

Except for the meager controls they already give you, the data flows only one way: from their db of history to your screen. You fencing off your interests would complicate this immeasurably. I would guess that it would raise the same problem that Google Search found with its domain blacklist feature 10+? years ago.

I don't think personal recommendations are a good use of developer effort, I think I'd be fine if it just showed me videos liked by other people who have liked the same videos I have.

Re: On YouTube’s recommendation system

#84
post #53

Why don't they let me tell them the interests that they are trying to infer? I'd be really happy to review sets of recommendations and indicate the kind of videos i like and the kind i don't.

In some vein they kind of do this by letting you like or dislike videos. The problem is almost all of them see this as an opportunity to signal others. Which means that your likes and dislikes are also value judgements, that you may or may not be comfortable making. They have at least have "not interested" but it's a clunky system. TikTok is having more luck because they're getting this data by watching WHEN you stop…

Huh, I'd assume they are all storing start, stop, and skips.

Re: On YouTube’s recommendation system

#85

For me, it seems like youtube's recommendation system mostly gets stuck in the local optimum. Once you watch a video from a particular channel, it'll keep on recommending the videos from the same channel for months. Soon, the recommendation just consists of stuff from the channels that have you already viewed. You rarely see new types of recommendations.

I've noticed this too. Occasionally I'll look up a how-to/DIY thing, watch one video for it, and then all my recommendations (even the ones on the front page) revolve around that topic. But it's not like I have a consistent interest in it, it was a one-off.

It's worse when it's decided that yeah, you are interested in that thing. My interests bounce between woodworking, homebrew vehicles, and strategy video games. The YT algorithm gets itself totally snarled when I watch a video from one side of that fence and assumes that that's all I want to see now. Watch a SuperfastMatt vehicle about the Jag he's retrofitting with the guts of a Tesla? Obviously you want seven more videos from his channel that you've already watched. And I like his channel! Not enough to watch the same seven videos again, though.

It also seems to have some classification issues, too. It lumps a lot of stuff into "DIY", and woodworking is part of it--but so is a lot of stuff that more fits under "construction" or "carpentry". No shade thrown, that stuff can be interesting too, but there's a large gulf between "crotchety woodworking dork going on about tablesaw safety" and "refinishing a backyard shed with glamour shots of a Home Depot sponsored miter saw set to bouncy stock music" that the algorithm does not seem to grok. Maybe it wants me to hatewatch that stuff, I dunno.

Re: On YouTube’s recommendation system

#86
post #8

For me, it seems like youtube's recommendation system mostly gets stuck in the local optimum. Once you watch a video from a particular channel, it'll keep on recommending the videos from the same channel for months. Soon, the recommendation just consists of stuff from the channels that have you already viewed. You rarely see new types of recommendations.

That mirrors my experience, too. In fact, I rarely find YouTube's recommendations to be useful. Occasionally, yes, but typically I just get that "more of the same" problem. However, that may be explained by how I use YouTube. The article is pretty vague and hand-wavy about how the recommendations really work, but it's possible that the fact that I don't drive my YouTube viewing from the recommended videos list makes…

I often search for “scientific documentary” or “variety” but the algo clearly doesn’t want me to discover new things, and doesn’t seem to include those searches to build the suggestions. It keeps repeating the 8 usual channels I’m bored of watching.

Re: On YouTube’s recommendation system

#88

> 3. Do recommendations drive viewers to increasingly extreme content? > As I’ve explained, we actively demote low-quality information in recommendations. But we also take the additional step of showing viewers authoritative videos about topics that may interest them. Say I watch a video about the COVID-19 vaccine. In my Up Next panel, I’ll see videos from reputable sources like Vox and Bloomberg Quicktake and won’t…

I'm sorry that you're in this painful situation. Personally I don't think the root problem is YouTube — rather, it's your parents' confirmation bias. They want to believe these things. I've watched Alex Jones videos myself. I did not leave wanting more and no recommendation algo could change that. (Lest I leave the wrong impression, the same is true of AOC going live on IG.)

This may be true. Ultimately my personal story is not worth much. I do know that there is a growing community of people who share this story, though. One does not personally need to be affected to notice an uptick in anti-knowledge behavior recently.

The greater point is that if X percent of the population has biases that if fed would lead them to deranged behavior, is it responsible to take advantage of that?

The reasonable discussion of course goes to how big of a number is X. If 100,000 families get destroyed per decade, no big deal in relation to those sweet advertising impressions, I guess. But if FB, YT, Reddit, etc collectively take the mostly hands off approach, does X become big enough to achieve history-book-level harm? Like making [effective representative democracy, pandemic response, ecologic collapse mitigation] impossible.

Re: On YouTube’s recommendation system

#89
post #11

The reccomendation system is absolutely dangerous. Just this summer, I watched one video on peppers out of curiousity. By the end of summer I've got multiple different varieties growing, and all the accessories to support them. I'm really enjoying my garden, but it's very unlikely I would have started one without the youtube algorithm. Similarly my doctor reccomended I exercise more, so I started running. Then I goog…

It's a consumption-driven economy, where those who are economically useless are driven out of the game.

Re: On YouTube’s recommendation system

#90

> For my oldest daughter, it was finding laughter and community with the Vlogbrothers. And for my oldest son, recommendations brought about a better understanding of linear algebra through animated explainers by 3Blue1Brown—with breaks to watch KSI videos. My first thought reading this was about reinforcing a stereotype as the son learning hard sciences and the daughter watching fun shows about people things. Did any…

It is reinforcing a stereotype.

But it is also telling his reality. It wouldn't be honest for him to say his daughter watches 3Blue1Brown if she doesn't, would it?

Post reply on HN