4 ms·

I'd like to see a serious investigation and discussion around the recommendation engines used to rank content added here. Something is seriously wrong with you

I'd like to see a serious investigation and discussion around the recommendation engines used to rank content added here. Something is seriously wrong with youtube's algorithm, whether in the technical implementation or how it deals with spam/unintended consequences. Kids watching cartoons get recommended messed up stuff (e.g. https://www.reddit.com/r/OutOfTheLoop/comments/65rkxv/what_is_up_with_the_weird_kids_videos_that_have/ https://www.reddit.com/r/OutOfTheLoop/comments/65rkxv/what_i... ), watch a single Jordan Peterson video and youtube will recommend you every fringe alt-right anti-feminism video it can find for months... It is toxic, causes real harm, normalizes extremist viewpoints and is shifting the Overton window in politics. I'm not advocating censorship in any way, my concern is if algorithms/implementation bias what people see towards what will get the strongest reaction it is at least in a similar class of harm as censorship... I have a hunch this is a natural consequence of using something like UCB (upper confidence bound) for contextual bandits to rank results when the feature vectors are using broad user / video clusters. Are explore vs exploit reinforcement learning algorithms used in production very susceptible to adversarial spam? Any evidence of fringe videos paying for viewbots etc...

bionoid 8y ago
> I'd like to see a serious investigation and discussion around the recommendation engines [...] Something is seriously wrong with youtube's algorithm I had a pretty surreal experience last year. My youtube use is almost exclusively for learning; coding, maths, astronomy, science, and so on. Recommendations were never great for me, usually recommending videos I've already watched. At some point I inadvertently clicked a link to a teenage girl's vlog, and watched maybe 5 seconds of it before closing. Instantly half of my recommendations were teenage girl's vlogs. I then misclicked a video in recommendations a few days later, and it turned into exclusively teenage girl vlogs. A year later, it's slightly recovered, but I still get them. I am pretty confident their recommendation would not be overrun with science if they inadvertantly click a video.
- empath75 8y ago
  I like to watch penn and teller fool us videos and if I click on one, my entire feed is full of magicians for weeks. They need to adjust how quickly they read signals. That said, I also watch numberphile and do get a bunch of like math suggestions like mathologer and v sauce.

bionoid 8y ago

> I'd like to see a serious investigation and discussion around the recommendation engines [...] Something is seriously wrong with youtube's algorithm I had a pretty surreal experience last year. My youtube use is almost exclusively for learning; coding, maths, astronomy, science, and so on. Recommendations were never great for me, usually recommending videos I've already watched. At some point I inadvertently clicked a link to a teenage girl's vlog, and watched maybe 5 seconds of it before closing. Instantly half of my recommendations were teenage girl's vlogs. I then misclicked a video in recommendations a few days later, and it turned into exclusively teenage girl vlogs. A year later, it's slightly recovered, but I still get them. I am pretty confident their recommendation would not be overrun with science if they inadvertantly click a video.

empath75 8y ago
I like to watch penn and teller fool us videos and if I click on one, my entire feed is full of magicians for weeks. They need to adjust how quickly they read signals. That said, I also watch numberphile and do get a bunch of like math suggestions like mathologer and v sauce.