4 ms·

Why TikTok in particular? What is the engineering story behind TikTok's recommendation system? How did they get it right?

keewee7 5y ago
TikTok seem to be learning from what the user is actually watching and for how long and not just the user's "Like"/"Not Interested In" actions. However it still seem to learn from the "Not Interested In" action more than any other platform.
- pedrosorio 5y ago
  This is a pretty misinformed take when it’s publicly known that YouTube was already doing this (learn from what the user is watching and for how long) the year Bytedance was founded (2012): https://blog.youtube/news-and-events/youtube-now-why-we-focus-on-watch-time/ https://blog.youtube/news-and-events/youtube-now-why-we-focu...
  - monkeybutton 5y ago
    Somehow they're doing it better. At least subjectively, people complain more about the YouTube algo's performance than tiktok. For the latter, the most common complaint is that it's too good.
    - leobg 5y ago
      TikToks are shorter. So in any given session, a user will give a lot more feedback to TikTok than to YouTube.
      - pedrosorio 5y ago
        Bingo
gammarator 5y ago
This article suggests that the “one video at a time” format makes attention metrics much more reliable: https://www.eugenewei.com/blog/2020/9/18/seeing-like-an-algorithm https://www.eugenewei.com/blog/2020/9/18/seeing-like-an-algo...
faizshah 5y ago
Not a direct answer to your question, but at least in English there isn’t that much stuff on how TikTok’s recommendation system works internally. This is the best breakdown I have found on TikTok’s recommendation system internals: https://leehanchung.github.io/2020-02-18-Tik-Tok-Algorithm/ https://leehanchung.github.io/2020-02-18-Tik-Tok-Algorithm/ There’s a more technical recent paper from bytedance as well: https://arxiv.org/pdf/2007.07203.pdf https://arxiv.org/pdf/2007.07203.pdf And another recent one on bytedance user profile system, this paper gives the deepest understanding of their recommendation system: https://www.cs.princeton.edu/courses/archive/spring21/cos598D/icde_2021_camera_ready.pdf https://www.cs.princeton.edu/courses/archive/spring21/cos598... There’s a good breakdown of the recent nytimes article on TikTok’s internals here: https://read.deeplearning.ai/the-batch/issue-122/ https://read.deeplearning.ai/the-batch/issue-122/ If anyone has found anything better let me know. If you try searching yourself you will want to try switching between the keywords “douyin” “toutiao” “bytedance” and “tiktok.”

keewee7 5y ago

TikTok seem to be learning from what the user is actually watching and for how long and not just the user's "Like"/"Not Interested In" actions. However it still seem to learn from the "Not Interested In" action more than any other platform.

pedrosorio 5y ago
This is a pretty misinformed take when it’s publicly known that YouTube was already doing this (learn from what the user is watching and for how long) the year Bytedance was founded (2012): https://blog.youtube/news-and-events/youtube-now-why-we-focus-on-watch-time/ https://blog.youtube/news-and-events/youtube-now-why-we-focu...
- monkeybutton 5y ago
  Somehow they're doing it better. At least subjectively, people complain more about the YouTube algo's performance than tiktok. For the latter, the most common complaint is that it's too good.
  - leobg 5y ago
    TikToks are shorter. So in any given session, a user will give a lot more feedback to TikTok than to YouTube.
    - pedrosorio 5y ago
      Bingo

gammarator 5y ago

This article suggests that the “one video at a time” format makes attention metrics much more reliable: https://www.eugenewei.com/blog/2020/9/18/seeing-like-an-algorithm https://www.eugenewei.com/blog/2020/9/18/seeing-like-an-algo...

faizshah 5y ago

Not a direct answer to your question, but at least in English there isn’t that much stuff on how TikTok’s recommendation system works internally. This is the best breakdown I have found on TikTok’s recommendation system internals: https://leehanchung.github.io/2020-02-18-Tik-Tok-Algorithm/ https://leehanchung.github.io/2020-02-18-Tik-Tok-Algorithm/ There’s a more technical recent paper from bytedance as well: https://arxiv.org/pdf/2007.07203.pdf https://arxiv.org/pdf/2007.07203.pdf And another recent one on bytedance user profile system, this paper gives the deepest understanding of their recommendation system: https://www.cs.princeton.edu/courses/archive/spring21/cos598D/icde_2021_camera_ready.pdf https://www.cs.princeton.edu/courses/archive/spring21/cos598... There’s a good breakdown of the recent nytimes article on TikTok’s internals here: https://read.deeplearning.ai/the-batch/issue-122/ https://read.deeplearning.ai/the-batch/issue-122/ If anyone has found anything better let me know. If you try searching yourself you will want to try switching between the keywords “douyin” “toutiao” “bytedance” and “tiktok.”