TikTok Reveals Details of How Its Algorithm Works (axios.com) 14
On a call with reporters Wednesday, TikTok executives said they were revealing details of their algorithm and data practices to dispel myths and rumors about the company. Axios reports: TikTok's algorithm uses machine learning to determine what content a user is most likely to engage with and serve them more of it, by finding videos that are similar or that are liked by people with similar user preferences. When users open TikTok for the first time, they are shown 8 popular videos featuring different trends, music, and topics. After that, the algorithm will continue to serve the user new iterations of 8 videos based on which videos the user engages with and what the user does. The algorithm identifies similar videos to those that have engaged a user based on video information, which could include details like captions, hashtags or sounds. Recommendations also take into account user device and account settings, which include data like language preference, country setting, and device type.
Once TikTok collects enough data about the user, the app is able to map a user's preferences in relation to similar users and group them into "clusters." Simultaneously, it also groups videos into "clusters" based on similar themes, like "basketball" or "bunnies." Using machine learning, the algorithm serves videos to users based on their proximity to other clusters of users and content that they like. TikTok's logic aims to avoid redundancies that could bore the user, like seeing multiple videos with the same music or from the same creator.
TikTok concedes that its ability to nail users' preferences so effectively means that its algorithm can produce "filter bubbles," reinforcing users' existing preferences rather than showing them more varied content, widening their horizons, or offering them opposing viewpoints. The company says that it's studying filter bubbles, including how long they last and how a user encounters them, to get better at breaking them when necessary. Since filter bubbles can reinforce conspiracy theories, hoaxes and other misinformation, TikTok's product and policy teams study which accounts and video information -- themes, hashtags, captions, and so on -- might be linked to misinformation. Videos or creators linked to misinformation are sent to the company's global content reviewers so they can be managed before they are distributed to users on the main feed, which is called the "For You" page.
Once TikTok collects enough data about the user, the app is able to map a user's preferences in relation to similar users and group them into "clusters." Simultaneously, it also groups videos into "clusters" based on similar themes, like "basketball" or "bunnies." Using machine learning, the algorithm serves videos to users based on their proximity to other clusters of users and content that they like. TikTok's logic aims to avoid redundancies that could bore the user, like seeing multiple videos with the same music or from the same creator.
TikTok concedes that its ability to nail users' preferences so effectively means that its algorithm can produce "filter bubbles," reinforcing users' existing preferences rather than showing them more varied content, widening their horizons, or offering them opposing viewpoints. The company says that it's studying filter bubbles, including how long they last and how a user encounters them, to get better at breaking them when necessary. Since filter bubbles can reinforce conspiracy theories, hoaxes and other misinformation, TikTok's product and policy teams study which accounts and video information -- themes, hashtags, captions, and so on -- might be linked to misinformation. Videos or creators linked to misinformation are sent to the company's global content reviewers so they can be managed before they are distributed to users on the main feed, which is called the "For You" page.
Re: (Score:1)
The bane of democracy.
Re: (Score:2)
They aren't saying everything. (Score:1)
Once TikTok collects enough data about the user, the app is able to map a user's preferences in relation to similar users and group them into "clusters."
What they aren't saying however is what other things are done with the data. Who else is provided access to the data? We already know that the CCP has unfettered access by law.
Re: (Score:1, Insightful)
I live in Canada and I don't intend to ever go to China. Also, Canada does not have a treaty of extradition with China. This means the Chinese government can't do anything against me and I absolutely don't care if the Chinese government can look into my life.
On the other hand, not only Canada has a treaty of extradition with the US, but it is a member of the Five Eyes alliance. Also, American companies have to give access to their data to the American government by law.
So between a product from a Chinese co
Re: (Score:1)
That's not actually true, the law only requires companies to hand over data on request. It's not like the NSA where they are just inside the system with full access to everything, all logged and preserved.
Of course it's possible that Chinese security has the same level of access as the NSA, likely even, but so far we have no evidence of that. We need a Chinese Snowden to leak it all.
Re: (Score:1)
Of course it's possible that Chinese security has the same level of access as the NSA, likely even, but so far we have no evidence of that. We need a Chinese Snowden to leak it all.
We don't need evidence to behave intelligently, which is to say, to assume that they could have the same level of access. It doesn't matter if it's true; if it could be true, we should assume that it is, and behave accordingly. Kind of like picking encryption key sizes... you have to plan for the future, not just the present.
Modpoints, fifty cents (Score:2)
Good work comrade
than why grap user's clipboards? (Score:1)
meta surveillance (Score:1)
What're actually wrong with TikTok (Score:1)
They are living in the reincarnation of the wrong era [wikipedia.org] with the politically (and ethnically) incorrect identity [wikipedia.org].
Can we get some input? (Score:3)
I'd like to see all media platform give users some options to feed into these algorithms. "Engage" is such a broad concept, it doesn't always mean that a person is seeing what they would like and we know of course that in many instances it means presenting them with controversial content that will simply get them worked up about something. Move beyond a binary "like" button or a simple "they watched it / didn't watch it" approach and let users tell the algorithm what direction to take. Then what they're providing can be more legitimately called a service that the user actually wants as opposed to something they use to manipulate. They could charge for it, make it a premium feature.
This never seemed to be the important part (Score:2)
Though the suggestion algorithm is good, it seems to have received an outsize proportion of admiration and commentary. I always reckoned their suggestion algorithm would be easy to match in performance for any serious play, and this article reinforces my opinion. (Snuck a machine learning term in there)
A less simply reproduced feature is their user interface for making new videos. Ask anyone who tried both TT and Instagram Reels (authored by a resource-rich outfit with many smart employees).